diff --git a/.cardboardlint.yml b/.cardboardlint.yml deleted file mode 100644 index 4a115a37cd..0000000000 --- a/.cardboardlint.yml +++ /dev/null @@ -1,5 +0,0 @@ -linters: -- pylint: - # pylintrc: pylintrc - filefilter: ['- test_*.py', '+ *.py', '- *.npy'] - # exclude: \ No newline at end of file diff --git a/.dockerignore b/.dockerignore index 8d8ad918c9..5b28aa99dc 100644 --- a/.dockerignore +++ b/.dockerignore @@ -6,4 +6,4 @@ TTS.egg-info/ tests/outputs/* tests/train_outputs/* __pycache__/ -*.pyc \ No newline at end of file +*.pyc diff --git a/.github/ISSUE_TEMPLATE/bug_report.yaml b/.github/ISSUE_TEMPLATE/bug_report.yaml index 34cde7e844..6a50c24562 100644 --- a/.github/ISSUE_TEMPLATE/bug_report.yaml +++ b/.github/ISSUE_TEMPLATE/bug_report.yaml @@ -59,7 +59,7 @@ body: You can either run `TTS/bin/collect_env_info.py` ```bash - wget https://raw.githubusercontent.com/coqui-ai/TTS/main/TTS/bin/collect_env_info.py + wget https://raw.githubusercontent.com/idiap/coqui-ai-TTS/main/TTS/bin/collect_env_info.py python collect_env_info.py ``` diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml index 05ca7db6bd..ccaaff7565 100644 --- a/.github/ISSUE_TEMPLATE/config.yml +++ b/.github/ISSUE_TEMPLATE/config.yml @@ -1,8 +1,8 @@ blank_issues_enabled: false contact_links: - name: CoquiTTS GitHub Discussions - url: https://github.com/coqui-ai/TTS/discussions + url: https://github.com/idiap/coqui-ai-TTS/discussions about: Please ask and answer questions here. - name: Coqui Security issue disclosure - url: mailto:info@coqui.ai + url: mailto:enno.hermann@gmail.com about: Please report security vulnerabilities here. diff --git a/.github/PR_TEMPLATE.md b/.github/PR_TEMPLATE.md index 330109c3bc..9e7605a4ef 100644 --- a/.github/PR_TEMPLATE.md +++ b/.github/PR_TEMPLATE.md @@ -5,11 +5,3 @@ Welcome to the 🐸TTS project! We are excited to see your interest, and appreci This repository is governed by the Contributor Covenant Code of Conduct. For more details, see the [CODE_OF_CONDUCT.md](CODE_OF_CONDUCT.md) file. In order to make a good pull request, please see our [CONTRIBUTING.md](CONTRIBUTING.md) file. - -Before accepting your pull request, you will be asked to sign a [Contributor License Agreement](https://cla-assistant.io/coqui-ai/TTS). - -This [Contributor License Agreement](https://cla-assistant.io/coqui-ai/TTS): - -- Protects you, Coqui, and the users of the code. -- Does not change your rights to use your contributions for any purpose. -- Does not change the license of the 🐸TTS project. It just makes the terms of your contribution clearer and lets us know you are OK to contribute. diff --git a/.github/actions/setup-uv/action.yml b/.github/actions/setup-uv/action.yml new file mode 100644 index 0000000000..c7dd4f5f99 --- /dev/null +++ b/.github/actions/setup-uv/action.yml @@ -0,0 +1,11 @@ +name: Setup uv + +runs: + using: 'composite' + steps: + - name: Install uv + uses: astral-sh/setup-uv@v4 + with: + version: "0.5.4" + enable-cache: true + cache-dependency-glob: "**/pyproject.toml" diff --git a/.github/stale.yml b/.github/stale.yml index e05eaf0b57..dd45bf098f 100644 --- a/.github/stale.yml +++ b/.github/stale.yml @@ -15,4 +15,3 @@ markComment: > for your contributions. You might also look our discussion channels. # Comment to post when closing a stale issue. Set to `false` to disable closeComment: false - diff --git a/.github/workflows/aux_tests.yml b/.github/workflows/aux_tests.yml deleted file mode 100644 index f4cb3ecfe1..0000000000 --- a/.github/workflows/aux_tests.yml +++ /dev/null @@ -1,51 +0,0 @@ -name: aux-tests - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y git make gcc - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Replace scarf urls - run: | - sed -i 's/https:\/\/coqui.gateway.scarf.sh\//https:\/\/github.com\/coqui-ai\/TTS\/releases\/download\//g' TTS/.models.json - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: make test_aux diff --git a/.github/workflows/data_tests.yml b/.github/workflows/data_tests.yml deleted file mode 100644 index 3d1e3f8c4d..0000000000 --- a/.github/workflows/data_tests.yml +++ /dev/null @@ -1,51 +0,0 @@ -name: data-tests - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y --no-install-recommends git make gcc - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Replace scarf urls - run: | - sed -i 's/https:\/\/coqui.gateway.scarf.sh\//https:\/\/github.com\/coqui-ai\/TTS\/releases\/download\//g' TTS/.models.json - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: make data_tests diff --git a/.github/workflows/docker.yaml b/.github/workflows/docker.yaml index 1f15159b42..249816a320 100644 --- a/.github/workflows/docker.yaml +++ b/.github/workflows/docker.yaml @@ -10,7 +10,7 @@ on: jobs: docker-build: name: "Build and push Docker image" - runs-on: ubuntu-20.04 + runs-on: ubuntu-latest strategy: matrix: arch: ["amd64"] @@ -18,7 +18,7 @@ jobs: - "nvidia/cuda:11.8.0-base-ubuntu22.04" # GPU enabled - "python:3.10.8-slim" # CPU only steps: - - uses: actions/checkout@v2 + - uses: actions/checkout@v4 - name: Log in to the Container registry uses: docker/login-action@v1 with: @@ -29,11 +29,11 @@ jobs: id: compute-tag run: | set -ex - base="ghcr.io/coqui-ai/tts" + base="ghcr.io/idiap/coqui-tts" tags="" # PR build if [[ ${{ matrix.base }} = "python:3.10.8-slim" ]]; then - base="ghcr.io/coqui-ai/tts-cpu" + base="ghcr.io/idiap/coqui-tts-cpu" fi if [[ "${{ startsWith(github.ref, 'refs/heads/') }}" = "true" ]]; then @@ -42,7 +42,7 @@ jobs: branch=${github_ref#*refs/heads/} # strip prefix to get branch name tags="${base}:${branch},${base}:${{ github.sha }}," elif [[ "${{ startsWith(github.ref, 'refs/tags/') }}" = "true" ]]; then - VERSION="v$(cat TTS/VERSION)" + VERSION="v$(grep -m 1 version pyproject.toml | grep -P '\d+\.\d+\.\d+' -o)" if [[ "${{ github.ref }}" != "refs/tags/${VERSION}" ]]; then echo "Pushed tag does not match VERSION file. Aborting push." exit 1 @@ -63,3 +63,58 @@ jobs: push: ${{ github.event_name == 'push' }} build-args: "BASE=${{ matrix.base }}" tags: ${{ steps.compute-tag.outputs.tags }} + docker-dev-build: + name: "Build the development Docker image" + runs-on: ubuntu-latest + strategy: + matrix: + arch: ["amd64"] + base: + - "nvidia/cuda:11.8.0-base-ubuntu22.04" # GPU enabled + steps: + - uses: actions/checkout@v4 + - name: Log in to the Container registry + uses: docker/login-action@v1 + with: + registry: ghcr.io + username: ${{ github.actor }} + password: ${{ secrets.GITHUB_TOKEN }} + - name: Compute Docker tags, check VERSION file matches tag + id: compute-tag + run: | + set -ex + base="ghcr.io/idiap/coqui-tts-dev" + tags="" # PR build + + if [[ ${{ matrix.base }} = "python:3.10.8-slim" ]]; then + base="ghcr.io/idiap/coqui-tts-dev-cpu" + fi + + if [[ "${{ startsWith(github.ref, 'refs/heads/') }}" = "true" ]]; then + # Push to branch + github_ref="${{ github.ref }}" + branch=${github_ref#*refs/heads/} # strip prefix to get branch name + tags="${base}:${branch},${base}:${{ github.sha }}," + elif [[ "${{ startsWith(github.ref, 'refs/tags/') }}" = "true" ]]; then + VERSION="v$(grep -m 1 version pyproject.toml | grep -P '\d+\.\d+\.\d+' -o)" + if [[ "${{ github.ref }}" != "refs/tags/${VERSION}" ]]; then + echo "Pushed tag does not match VERSION file. Aborting push." + exit 1 + fi + tags="${base}:${VERSION},${base}:latest,${base}:${{ github.sha }}" + fi + echo "::set-output name=tags::${tags}" + - name: Set up QEMU + uses: docker/setup-qemu-action@v1 + - name: Set up Docker Buildx + id: buildx + uses: docker/setup-buildx-action@v1 + - name: Build and push + uses: docker/build-push-action@v2 + with: + context: . + file: dockerfiles/Dockerfile.dev + platforms: linux/${{ matrix.arch }} + push: false + build-args: "BASE=${{ matrix.base }}" + tags: ${{ steps.compute-tag.outputs.tags }} diff --git a/.github/workflows/inference_tests.yml b/.github/workflows/inference_tests.yml deleted file mode 100644 index d2159027b6..0000000000 --- a/.github/workflows/inference_tests.yml +++ /dev/null @@ -1,53 +0,0 @@ -name: inference_tests - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: | - export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y --no-install-recommends git make gcc - sudo apt-get install espeak-ng - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Replace scarf urls - run: | - sed -i 's/https:\/\/coqui.gateway.scarf.sh\//https:\/\/github.com\/coqui-ai\/TTS\/releases\/download\//g' TTS/.models.json - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: make inference_tests diff --git a/.github/workflows/pypi-release.yml b/.github/workflows/pypi-release.yml index 2bbcf3cd70..ef74c60da6 100644 --- a/.github/workflows/pypi-release.yml +++ b/.github/workflows/pypi-release.yml @@ -7,88 +7,48 @@ defaults: shell: bash jobs: - build-sdist: - runs-on: ubuntu-20.04 + build: + runs-on: ubuntu-latest steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 + - name: Setup uv + uses: ./.github/actions/setup-uv - name: Verify tag matches version run: | set -ex - version=$(cat TTS/VERSION) + version=$(grep -m 1 version pyproject.toml | grep -P '\d+\.\d+\.\d+' -o) tag="${GITHUB_REF/refs\/tags\/}" if [[ "v$version" != "$tag" ]]; then exit 1 fi - - uses: actions/setup-python@v2 - with: - python-version: 3.9 - - run: | - python -m pip install -U pip setuptools wheel build - - run: | - python -m build - - run: | - pip install dist/*.tar.gz - - uses: actions/upload-artifact@v2 - with: - name: sdist - path: dist/*.tar.gz - build-wheels: - runs-on: ubuntu-20.04 - strategy: - matrix: - python-version: ["3.9", "3.10", "3.11"] - steps: - - uses: actions/checkout@v3 - - uses: actions/setup-python@v2 - with: - python-version: ${{ matrix.python-version }} - - name: Install pip requirements - run: | - python -m pip install -U pip setuptools wheel build - python -m pip install -r requirements.txt - - name: Setup and install manylinux1_x86_64 wheel + - name: Set up Python + run: uv python install 3.12 + - name: Build sdist and wheel + run: uv build + - name: Test installation of sdist and wheel run: | - python setup.py bdist_wheel --plat-name=manylinux1_x86_64 - python -m pip install dist/*-manylinux*.whl - - uses: actions/upload-artifact@v2 + uv venv --no-project + uv pip install dist/*.tar.gz + uv pip install dist/*.whl + - uses: actions/upload-artifact@v4 with: - name: wheel-${{ matrix.python-version }} - path: dist/*-manylinux*.whl + name: build + path: dist/* publish-artifacts: - runs-on: ubuntu-20.04 - needs: [build-sdist, build-wheels] + name: Publish to PyPI + runs-on: ubuntu-latest + needs: [build] + environment: + name: release + url: https://pypi.org/p/coqui-tts + permissions: + id-token: write steps: - - run: | - mkdir dist - - uses: actions/download-artifact@v2 + - uses: actions/download-artifact@v4 with: - name: "sdist" - path: "dist/" - - uses: actions/download-artifact@v2 - with: - name: "wheel-3.9" - path: "dist/" - - uses: actions/download-artifact@v2 - with: - name: "wheel-3.10" - path: "dist/" - - uses: actions/download-artifact@v2 - with: - name: "wheel-3.11" path: "dist/" + name: build - run: | ls -lh dist/ - - name: Setup PyPI config - run: | - cat << EOF > ~/.pypirc - [pypi] - username=__token__ - password=${{ secrets.PYPI_TOKEN }} - EOF - - uses: actions/setup-python@v2 - with: - python-version: 3.9 - - run: | - python -m pip install twine - - run: | - twine upload --repository pypi dist/* + - name: Publish package distributions to PyPI + uses: pypa/gh-action-pypi-publish@release/v1 diff --git a/.github/workflows/style_check.yml b/.github/workflows/style_check.yml index b7c6393baa..d1060f6be2 100644 --- a/.github/workflows/style_check.yml +++ b/.github/workflows/style_check.yml @@ -7,40 +7,17 @@ on: pull_request: types: [opened, synchronize, reopened] jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: + lint: runs-on: ubuntu-latest strategy: fail-fast: false matrix: python-version: [3.9] - experimental: [false] steps: - - uses: actions/checkout@v3 + - uses: actions/checkout@v4 + - name: Setup uv + uses: ./.github/actions/setup-uv - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y git make gcc - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Style check - run: make style + run: uv python install ${{ matrix.python-version }} + - name: Lint check + run: make lint diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml new file mode 100644 index 0000000000..8d639d5dee --- /dev/null +++ b/.github/workflows/tests.yml @@ -0,0 +1,127 @@ +name: test + +on: + push: + branches: + - main + pull_request: + types: [opened, synchronize, reopened] + workflow_dispatch: + inputs: + trainer_branch: + description: "Branch of Trainer to test" + required: false + default: "main" + coqpit_branch: + description: "Branch of Coqpit to test" + required: false + default: "main" +jobs: + unit: + runs-on: ubuntu-latest + strategy: + fail-fast: false + matrix: + python-version: [3.9, "3.10", "3.11", "3.12"] + subset: ["data_tests", "inference_tests", "test_aux", "test_text"] + steps: + - uses: actions/checkout@v4 + - name: Setup uv + uses: ./.github/actions/setup-uv + - name: Set up Python ${{ matrix.python-version }} + run: uv python install ${{ matrix.python-version }} + - name: Install Espeak + if: contains(fromJSON('["inference_tests", "test_text"]'), matrix.subset) + run: | + sudo apt-get update + sudo apt-get install espeak espeak-ng + - name: Install dependencies + run: | + sudo apt-get update + sudo apt-get install -y --no-install-recommends git make gcc + make system-deps + - name: Install custom Trainer and/or Coqpit if requested + run: | + if [[ -n "${{ github.event.inputs.trainer_branch }}" ]]; then + uv add git+https://github.com/idiap/coqui-ai-Trainer --branch ${{ github.event.inputs.trainer_branch }} + fi + if [[ -n "${{ github.event.inputs.coqpit_branch }}" ]]; then + uv add git+https://github.com/idiap/coqui-ai-coqpit --branch ${{ github.event.inputs.coqpit_branch }} + fi + - name: Unit tests + run: | + resolution=highest + if [ "${{ matrix.python-version }}" == "3.9" ]; then + resolution=lowest-direct + fi + uv run --resolution=$resolution --extra server --extra languages make ${{ matrix.subset }} + - name: Upload coverage data + uses: actions/upload-artifact@v4 + with: + include-hidden-files: true + name: coverage-data-${{ matrix.subset }}-${{ matrix.python-version }} + path: .coverage.* + if-no-files-found: ignore + integration: + runs-on: ubuntu-latest + strategy: + fail-fast: false + matrix: + python-version: ["3.9", "3.12"] + subset: ["test_tts", "test_tts2", "test_vocoder", "test_xtts", "test_zoo0", "test_zoo1", "test_zoo2"] + steps: + - uses: actions/checkout@v4 + - name: Setup uv + uses: ./.github/actions/setup-uv + - name: Set up Python ${{ matrix.python-version }} + run: uv python install ${{ matrix.python-version }} + - name: Install Espeak + if: contains(fromJSON('["test_tts", "test_tts2", "test_xtts", "test_zoo0", "test_zoo1", "test_zoo2"]'), matrix.subset) + run: | + sudo apt-get update + sudo apt-get install espeak espeak-ng + - name: Install dependencies + run: | + sudo apt-get update + sudo apt-get install -y --no-install-recommends git make gcc + make system-deps + - name: Install custom Trainer and/or Coqpit if requested + run: | + if [[ -n "${{ github.event.inputs.trainer_branch }}" ]]; then + uv add git+https://github.com/idiap/coqui-ai-Trainer --branch ${{ github.event.inputs.trainer_branch }} + fi + if [[ -n "${{ github.event.inputs.coqpit_branch }}" ]]; then + uv add git+https://github.com/idiap/coqui-ai-coqpit --branch ${{ github.event.inputs.coqpit_branch }} + fi + - name: Integration tests + run: | + resolution=highest + if [ "${{ matrix.python-version }}" == "3.9" ]; then + resolution=lowest-direct + fi + uv run --resolution=$resolution --extra server --extra languages make ${{ matrix.subset }} + - name: Upload coverage data + uses: actions/upload-artifact@v4 + with: + include-hidden-files: true + name: coverage-data-${{ matrix.subset }}-${{ matrix.python-version }} + path: .coverage.* + if-no-files-found: ignore + coverage: + if: always() + needs: [unit, integration] + runs-on: ubuntu-latest + steps: + - uses: actions/checkout@v4 + - name: Setup uv + uses: ./.github/actions/setup-uv + - uses: actions/download-artifact@v4 + with: + pattern: coverage-data-* + merge-multiple: true + - name: Combine coverage + run: | + uv python install + uvx coverage combine + uvx coverage html --skip-covered --skip-empty + uvx coverage report --format=markdown >> $GITHUB_STEP_SUMMARY diff --git a/.github/workflows/text_tests.yml b/.github/workflows/text_tests.yml deleted file mode 100644 index 78d3026d7f..0000000000 --- a/.github/workflows/text_tests.yml +++ /dev/null @@ -1,50 +0,0 @@ -name: text-tests - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y --no-install-recommends git make gcc - sudo apt-get install espeak - sudo apt-get install espeak-ng - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: make test_text diff --git a/.github/workflows/tts_tests.yml b/.github/workflows/tts_tests.yml deleted file mode 100644 index 5074cded6d..0000000000 --- a/.github/workflows/tts_tests.yml +++ /dev/null @@ -1,53 +0,0 @@ -name: tts-tests - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y --no-install-recommends git make gcc - sudo apt-get install espeak - sudo apt-get install espeak-ng - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Replace scarf urls - run: | - sed -i 's/https:\/\/coqui.gateway.scarf.sh\//https:\/\/github.com\/coqui-ai\/TTS\/releases\/download\//g' TTS/.models.json - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: make test_tts diff --git a/.github/workflows/tts_tests2.yml b/.github/workflows/tts_tests2.yml deleted file mode 100644 index f64433f8df..0000000000 --- a/.github/workflows/tts_tests2.yml +++ /dev/null @@ -1,53 +0,0 @@ -name: tts-tests2 - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y --no-install-recommends git make gcc - sudo apt-get install espeak - sudo apt-get install espeak-ng - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Replace scarf urls - run: | - sed -i 's/https:\/\/coqui.gateway.scarf.sh\//https:\/\/github.com\/coqui-ai\/TTS\/releases\/download\//g' TTS/.models.json - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: make test_tts2 diff --git a/.github/workflows/vocoder_tests.yml b/.github/workflows/vocoder_tests.yml deleted file mode 100644 index 6519ee3fef..0000000000 --- a/.github/workflows/vocoder_tests.yml +++ /dev/null @@ -1,48 +0,0 @@ -name: vocoder-tests - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y git make gcc - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: make test_vocoder diff --git a/.github/workflows/xtts_tests.yml b/.github/workflows/xtts_tests.yml deleted file mode 100644 index be367f3547..0000000000 --- a/.github/workflows/xtts_tests.yml +++ /dev/null @@ -1,53 +0,0 @@ -name: xtts-tests - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y --no-install-recommends git make gcc - sudo apt-get install espeak - sudo apt-get install espeak-ng - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Replace scarf urls - run: | - sed -i 's/https:\/\/coqui.gateway.scarf.sh\//https:\/\/github.com\/coqui-ai\/TTS\/releases\/download\//g' TTS/.models.json - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: make test_xtts diff --git a/.github/workflows/zoo_tests0.yml b/.github/workflows/zoo_tests0.yml deleted file mode 100644 index 13f47a938b..0000000000 --- a/.github/workflows/zoo_tests0.yml +++ /dev/null @@ -1,54 +0,0 @@ -name: zoo-tests-0 - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y git make gcc - sudo apt-get install espeak espeak-ng - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Replace scarf urls - run: | - sed -i 's/https:\/\/coqui.gateway.scarf.sh\//https:\/\/github.com\/coqui-ai\/TTS\/releases\/download\//g' TTS/.models.json - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: | - nose2 -F -v -B TTS tests.zoo_tests.test_models.test_models_offset_0_step_3 - nose2 -F -v -B TTS tests.zoo_tests.test_models.test_voice_conversion diff --git a/.github/workflows/zoo_tests1.yml b/.github/workflows/zoo_tests1.yml deleted file mode 100644 index 00f13397fa..0000000000 --- a/.github/workflows/zoo_tests1.yml +++ /dev/null @@ -1,53 +0,0 @@ -name: zoo-tests-1 - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y git make gcc - sudo apt-get install espeak espeak-ng - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Replace scarf urls - run: | - sed -i 's/https:\/\/coqui.gateway.scarf.sh\/hf\/bark\//https:\/\/huggingface.co\/erogol\/bark\/resolve\/main\//g' TTS/.models.json - sed -i 's/https:\/\/coqui.gateway.scarf.sh\//https:\/\/github.com\/coqui-ai\/TTS\/releases\/download\//g' TTS/.models.json - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: nose2 -F -v -B --with-coverage --coverage TTS tests.zoo_tests.test_models.test_models_offset_1_step_3 diff --git a/.github/workflows/zoo_tests2.yml b/.github/workflows/zoo_tests2.yml deleted file mode 100644 index 310a831a8b..0000000000 --- a/.github/workflows/zoo_tests2.yml +++ /dev/null @@ -1,52 +0,0 @@ -name: zoo-tests-2 - -on: - push: - branches: - - main - pull_request: - types: [opened, synchronize, reopened] -jobs: - check_skip: - runs-on: ubuntu-latest - if: "! contains(github.event.head_commit.message, '[ci skip]')" - steps: - - run: echo "${{ github.event.head_commit.message }}" - - test: - runs-on: ubuntu-latest - strategy: - fail-fast: false - matrix: - python-version: [3.9, "3.10", "3.11"] - experimental: [false] - steps: - - uses: actions/checkout@v3 - - name: Set up Python ${{ matrix.python-version }} - uses: actions/setup-python@v4 - with: - python-version: ${{ matrix.python-version }} - architecture: x64 - cache: 'pip' - cache-dependency-path: 'requirements*' - - name: check OS - run: cat /etc/os-release - - name: set ENV - run: export TRAINER_TELEMETRY=0 - - name: Install dependencies - run: | - sudo apt-get update - sudo apt-get install -y git make gcc - sudo apt-get install espeak espeak-ng - make system-deps - - name: Install/upgrade Python setup deps - run: python3 -m pip install --upgrade pip setuptools wheel - - name: Replace scarf urls - run: | - sed -i 's/https:\/\/coqui.gateway.scarf.sh\//https:\/\/github.com\/coqui-ai\/TTS\/releases\/download\//g' TTS/.models.json - - name: Install TTS - run: | - python3 -m pip install .[all] - python3 setup.py egg_info - - name: Unit tests - run: nose2 -F -v -B --with-coverage --coverage TTS tests.zoo_tests.test_models.test_models_offset_2_step_3 diff --git a/.gitignore b/.gitignore index 22ec6e410a..d9f992275c 100644 --- a/.gitignore +++ b/.gitignore @@ -1,3 +1,5 @@ +uv.lock + WadaSNR/ .idea/ *.pyc @@ -169,4 +171,4 @@ wandb depot/* coqui_recipes/* local_scripts/* -coqui_demos/* \ No newline at end of file +coqui_demos/* diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index 911f2a838e..62420e9958 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -1,27 +1,19 @@ repos: - - repo: 'https://github.com/pre-commit/pre-commit-hooks' - rev: v2.3.0 + - repo: "https://github.com/pre-commit/pre-commit-hooks" + rev: v5.0.0 hooks: + - id: check-json + files: "TTS/.models.json" - id: check-yaml - id: end-of-file-fixer - id: trailing-whitespace - - repo: 'https://github.com/psf/black' - rev: 22.3.0 + - repo: "https://github.com/psf/black" + rev: 24.2.0 hooks: - id: black language_version: python3 - - repo: https://github.com/pycqa/isort - rev: 5.8.0 + - repo: https://github.com/astral-sh/ruff-pre-commit + rev: v0.7.0 hooks: - - id: isort - name: isort (python) - - id: isort - name: isort (cython) - types: [cython] - - id: isort - name: isort (pyi) - types: [pyi] - - repo: https://github.com/pycqa/pylint - rev: v2.8.2 - hooks: - - id: pylint + - id: ruff + args: [--fix, --exit-non-zero-on-fix] diff --git a/.pylintrc b/.pylintrc deleted file mode 100644 index 49a9dbdd2c..0000000000 --- a/.pylintrc +++ /dev/null @@ -1,599 +0,0 @@ -[MASTER] - -# A comma-separated list of package or module names from where C extensions may -# be loaded. Extensions are loading into the active Python interpreter and may -# run arbitrary code. -extension-pkg-whitelist= - -# Add files or directories to the blacklist. They should be base names, not -# paths. -ignore=CVS - -# Add files or directories matching the regex patterns to the blacklist. The -# regex matches against base names, not paths. -ignore-patterns= - -# Python code to execute, usually for sys.path manipulation such as -# pygtk.require(). -#init-hook= - -# Use multiple processes to speed up Pylint. Specifying 0 will auto-detect the -# number of processors available to use. -jobs=1 - -# Control the amount of potential inferred values when inferring a single -# object. This can help the performance when dealing with large functions or -# complex, nested conditions. -limit-inference-results=100 - -# List of plugins (as comma separated values of python modules names) to load, -# usually to register additional checkers. -load-plugins= - -# Pickle collected data for later comparisons. -persistent=yes - -# Specify a configuration file. -#rcfile= - -# When enabled, pylint would attempt to guess common misconfiguration and emit -# user-friendly hints instead of false-positive error messages. -suggestion-mode=yes - -# Allow loading of arbitrary C extensions. Extensions are imported into the -# active Python interpreter and may run arbitrary code. -unsafe-load-any-extension=no - - -[MESSAGES CONTROL] - -# Only show warnings with the listed confidence levels. Leave empty to show -# all. Valid levels: HIGH, INFERENCE, INFERENCE_FAILURE, UNDEFINED. -confidence= - -# Disable the message, report, category or checker with the given id(s). You -# can either give multiple identifiers separated by comma (,) or put this -# option multiple times (only on the command line, not in the configuration -# file where it should appear only once). You can also use "--disable=all" to -# disable everything first and then reenable specific checks. For example, if -# you want to run only the similarities checker, you can use "--disable=all -# --enable=similarities". If you want to run only the classes checker, but have -# no Warning level messages displayed, use "--disable=all --enable=classes -# --disable=W". -disable=missing-docstring, - too-many-public-methods, - too-many-lines, - bare-except, - ## for avoiding weird p3.6 CI linter error - ## TODO: see later if we can remove this - assigning-non-slot, - unsupported-assignment-operation, - ## end - line-too-long, - fixme, - wrong-import-order, - ungrouped-imports, - wrong-import-position, - import-error, - invalid-name, - too-many-instance-attributes, - arguments-differ, - arguments-renamed, - no-name-in-module, - no-member, - unsubscriptable-object, - print-statement, - parameter-unpacking, - unpacking-in-except, - old-raise-syntax, - backtick, - long-suffix, - old-ne-operator, - old-octal-literal, - import-star-module-level, - non-ascii-bytes-literal, - raw-checker-failed, - bad-inline-option, - locally-disabled, - file-ignored, - suppressed-message, - useless-suppression, - deprecated-pragma, - use-symbolic-message-instead, - useless-object-inheritance, - too-few-public-methods, - too-many-branches, - too-many-arguments, - too-many-locals, - too-many-statements, - apply-builtin, - basestring-builtin, - buffer-builtin, - cmp-builtin, - coerce-builtin, - execfile-builtin, - file-builtin, - long-builtin, - raw_input-builtin, - reduce-builtin, - standarderror-builtin, - unicode-builtin, - xrange-builtin, - coerce-method, - delslice-method, - getslice-method, - setslice-method, - no-absolute-import, - old-division, - dict-iter-method, - dict-view-method, - next-method-called, - metaclass-assignment, - indexing-exception, - raising-string, - reload-builtin, - oct-method, - hex-method, - nonzero-method, - cmp-method, - input-builtin, - round-builtin, - intern-builtin, - unichr-builtin, - map-builtin-not-iterating, - zip-builtin-not-iterating, - range-builtin-not-iterating, - filter-builtin-not-iterating, - using-cmp-argument, - eq-without-hash, - div-method, - idiv-method, - rdiv-method, - exception-message-attribute, - invalid-str-codec, - sys-max-int, - bad-python3-import, - deprecated-string-function, - deprecated-str-translate-call, - deprecated-itertools-function, - deprecated-types-field, - next-method-defined, - dict-items-not-iterating, - dict-keys-not-iterating, - dict-values-not-iterating, - deprecated-operator-function, - deprecated-urllib-function, - xreadlines-attribute, - deprecated-sys-function, - exception-escape, - comprehension-escape, - duplicate-code, - not-callable, - import-outside-toplevel, - logging-fstring-interpolation, - logging-not-lazy - -# Enable the message, report, category or checker with the given id(s). You can -# either give multiple identifier separated by comma (,) or put this option -# multiple time (only on the command line, not in the configuration file where -# it should appear only once). See also the "--disable" option for examples. -enable=c-extension-no-member - - -[REPORTS] - -# Python expression which should return a note less than 10 (10 is the highest -# note). You have access to the variables errors warning, statement which -# respectively contain the number of errors / warnings messages and the total -# number of statements analyzed. This is used by the global evaluation report -# (RP0004). -evaluation=10.0 - ((float(5 * error + warning + refactor + convention) / statement) * 10) - -# Template used to display messages. This is a python new-style format string -# used to format the message information. See doc for all details. -#msg-template= - -# Set the output format. Available formats are text, parseable, colorized, json -# and msvs (visual studio). You can also give a reporter class, e.g. -# mypackage.mymodule.MyReporterClass. -output-format=text - -# Tells whether to display a full report or only the messages. -reports=no - -# Activate the evaluation score. -score=yes - - -[REFACTORING] - -# Maximum number of nested blocks for function / method body -max-nested-blocks=5 - -# Complete name of functions that never returns. When checking for -# inconsistent-return-statements if a never returning function is called then -# it will be considered as an explicit return statement and no message will be -# printed. -never-returning-functions=sys.exit - - -[LOGGING] - -# Format style used to check logging format string. `old` means using % -# formatting, while `new` is for `{}` formatting. -logging-format-style=old - -# Logging modules to check that the string format arguments are in logging -# function parameter format. -logging-modules=logging - - -[SPELLING] - -# Limits count of emitted suggestions for spelling mistakes. -max-spelling-suggestions=4 - -# Spelling dictionary name. Available dictionaries: none. To make it working -# install python-enchant package.. -spelling-dict= - -# List of comma separated words that should not be checked. -spelling-ignore-words= - -# A path to a file that contains private dictionary; one word per line. -spelling-private-dict-file= - -# Tells whether to store unknown words to indicated private dictionary in -# --spelling-private-dict-file option instead of raising a message. -spelling-store-unknown-words=no - - -[MISCELLANEOUS] - -# List of note tags to take in consideration, separated by a comma. -notes=FIXME, - XXX, - TODO - - -[TYPECHECK] - -# List of decorators that produce context managers, such as -# contextlib.contextmanager. Add to this list to register other decorators that -# produce valid context managers. -contextmanager-decorators=contextlib.contextmanager - -# List of members which are set dynamically and missed by pylint inference -# system, and so shouldn't trigger E1101 when accessed. Python regular -# expressions are accepted. -generated-members=numpy.*,torch.* - -# Tells whether missing members accessed in mixin class should be ignored. A -# mixin class is detected if its name ends with "mixin" (case insensitive). -ignore-mixin-members=yes - -# Tells whether to warn about missing members when the owner of the attribute -# is inferred to be None. -ignore-none=yes - -# This flag controls whether pylint should warn about no-member and similar -# checks whenever an opaque object is returned when inferring. The inference -# can return multiple potential results while evaluating a Python object, but -# some branches might not be evaluated, which results in partial inference. In -# that case, it might be useful to still emit no-member and other checks for -# the rest of the inferred objects. -ignore-on-opaque-inference=yes - -# List of class names for which member attributes should not be checked (useful -# for classes with dynamically set attributes). This supports the use of -# qualified names. -ignored-classes=optparse.Values,thread._local,_thread._local - -# List of module names for which member attributes should not be checked -# (useful for modules/projects where namespaces are manipulated during runtime -# and thus existing member attributes cannot be deduced by static analysis. It -# supports qualified module names, as well as Unix pattern matching. -ignored-modules= - -# Show a hint with possible names when a member name was not found. The aspect -# of finding the hint is based on edit distance. -missing-member-hint=yes - -# The minimum edit distance a name should have in order to be considered a -# similar match for a missing member name. -missing-member-hint-distance=1 - -# The total number of similar names that should be taken in consideration when -# showing a hint for a missing member. -missing-member-max-choices=1 - - -[VARIABLES] - -# List of additional names supposed to be defined in builtins. Remember that -# you should avoid defining new builtins when possible. -additional-builtins= - -# Tells whether unused global variables should be treated as a violation. -allow-global-unused-variables=yes - -# List of strings which can identify a callback function by name. A callback -# name must start or end with one of those strings. -callbacks=cb_, - _cb - -# A regular expression matching the name of dummy variables (i.e. expected to -# not be used). -dummy-variables-rgx=_+$|(_[a-zA-Z0-9_]*[a-zA-Z0-9]+?$)|dummy|^ignored_|^unused_ - -# Argument names that match this expression will be ignored. Default to name -# with leading underscore. -ignored-argument-names=_.*|^ignored_|^unused_ - -# Tells whether we should check for unused import in __init__ files. -init-import=no - -# List of qualified module names which can have objects that can redefine -# builtins. -redefining-builtins-modules=six.moves,past.builtins,future.builtins,builtins,io - - -[FORMAT] - -# Expected format of line ending, e.g. empty (any line ending), LF or CRLF. -expected-line-ending-format= - -# Regexp for a line that is allowed to be longer than the limit. -ignore-long-lines=^\s*(# )??$ - -# Number of spaces of indent required inside a hanging or continued line. -indent-after-paren=4 - -# String used as indentation unit. This is usually " " (4 spaces) or "\t" (1 -# tab). -indent-string=' ' - -# Maximum number of characters on a single line. -max-line-length=120 - -# Maximum number of lines in a module. -max-module-lines=1000 - -# List of optional constructs for which whitespace checking is disabled. `dict- -# separator` is used to allow tabulation in dicts, etc.: {1 : 1,\n222: 2}. -# `trailing-comma` allows a space between comma and closing bracket: (a, ). -# `empty-line` allows space-only lines. -no-space-check=trailing-comma, - dict-separator - -# Allow the body of a class to be on the same line as the declaration if body -# contains single statement. -single-line-class-stmt=no - -# Allow the body of an if to be on the same line as the test if there is no -# else. -single-line-if-stmt=no - - -[SIMILARITIES] - -# Ignore comments when computing similarities. -ignore-comments=yes - -# Ignore docstrings when computing similarities. -ignore-docstrings=yes - -# Ignore imports when computing similarities. -ignore-imports=no - -# Minimum lines number of a similarity. -min-similarity-lines=4 - - -[BASIC] - -# Naming style matching correct argument names. -argument-naming-style=snake_case - -# Regular expression matching correct argument names. Overrides argument- -# naming-style. -argument-rgx=[a-z_][a-z0-9_]{0,30}$ - -# Naming style matching correct attribute names. -attr-naming-style=snake_case - -# Regular expression matching correct attribute names. Overrides attr-naming- -# style. -#attr-rgx= - -# Bad variable names which should always be refused, separated by a comma. -bad-names= - -# Naming style matching correct class attribute names. -class-attribute-naming-style=any - -# Regular expression matching correct class attribute names. Overrides class- -# attribute-naming-style. -#class-attribute-rgx= - -# Naming style matching correct class names. -class-naming-style=PascalCase - -# Regular expression matching correct class names. Overrides class-naming- -# style. -#class-rgx= - -# Naming style matching correct constant names. -const-naming-style=UPPER_CASE - -# Regular expression matching correct constant names. Overrides const-naming- -# style. -#const-rgx= - -# Minimum line length for functions/classes that require docstrings, shorter -# ones are exempt. -docstring-min-length=-1 - -# Naming style matching correct function names. -function-naming-style=snake_case - -# Regular expression matching correct function names. Overrides function- -# naming-style. -#function-rgx= - -# Good variable names which should always be accepted, separated by a comma. -good-names=i, - j, - k, - x, - ex, - Run, - _ - -# Include a hint for the correct naming format with invalid-name. -include-naming-hint=no - -# Naming style matching correct inline iteration names. -inlinevar-naming-style=any - -# Regular expression matching correct inline iteration names. Overrides -# inlinevar-naming-style. -#inlinevar-rgx= - -# Naming style matching correct method names. -method-naming-style=snake_case - -# Regular expression matching correct method names. Overrides method-naming- -# style. -#method-rgx= - -# Naming style matching correct module names. -module-naming-style=snake_case - -# Regular expression matching correct module names. Overrides module-naming- -# style. -#module-rgx= - -# Colon-delimited sets of names that determine each other's naming style when -# the name regexes allow several styles. -name-group= - -# Regular expression which should only match function or class names that do -# not require a docstring. -no-docstring-rgx=^_ - -# List of decorators that produce properties, such as abc.abstractproperty. Add -# to this list to register other decorators that produce valid properties. -# These decorators are taken in consideration only for invalid-name. -property-classes=abc.abstractproperty - -# Naming style matching correct variable names. -variable-naming-style=snake_case - -# Regular expression matching correct variable names. Overrides variable- -# naming-style. -variable-rgx=[a-z_][a-z0-9_]{0,30}$ - - -[STRING] - -# This flag controls whether the implicit-str-concat-in-sequence should -# generate a warning on implicit string concatenation in sequences defined over -# several lines. -check-str-concat-over-line-jumps=no - - -[IMPORTS] - -# Allow wildcard imports from modules that define __all__. -allow-wildcard-with-all=no - -# Analyse import fallback blocks. This can be used to support both Python 2 and -# 3 compatible code, which means that the block might have code that exists -# only in one or another interpreter, leading to false positives when analysed. -analyse-fallback-blocks=no - -# Deprecated modules which should not be used, separated by a comma. -deprecated-modules=optparse,tkinter.tix - -# Create a graph of external dependencies in the given file (report RP0402 must -# not be disabled). -ext-import-graph= - -# Create a graph of every (i.e. internal and external) dependencies in the -# given file (report RP0402 must not be disabled). -import-graph= - -# Create a graph of internal dependencies in the given file (report RP0402 must -# not be disabled). -int-import-graph= - -# Force import order to recognize a module as part of the standard -# compatibility libraries. -known-standard-library= - -# Force import order to recognize a module as part of a third party library. -known-third-party=enchant - - -[CLASSES] - -# List of method names used to declare (i.e. assign) instance attributes. -defining-attr-methods=__init__, - __new__, - setUp - -# List of member names, which should be excluded from the protected access -# warning. -exclude-protected=_asdict, - _fields, - _replace, - _source, - _make - -# List of valid names for the first argument in a class method. -valid-classmethod-first-arg=cls - -# List of valid names for the first argument in a metaclass class method. -valid-metaclass-classmethod-first-arg=cls - - -[DESIGN] - -# Maximum number of arguments for function / method. -max-args=5 - -# Maximum number of attributes for a class (see R0902). -max-attributes=7 - -# Maximum number of boolean expressions in an if statement. -max-bool-expr=5 - -# Maximum number of branch for function / method body. -max-branches=12 - -# Maximum number of locals for function / method body. -max-locals=15 - -# Maximum number of parents for a class (see R0901). -max-parents=15 - -# Maximum number of public methods for a class (see R0904). -max-public-methods=20 - -# Maximum number of return / yield for function / method body. -max-returns=6 - -# Maximum number of statements in function / method body. -max-statements=50 - -# Minimum number of public methods for a class (see R0903). -min-public-methods=2 - - -[EXCEPTIONS] - -# Exceptions that will emit a warning when being caught. Defaults to -# "BaseException, Exception". -overgeneral-exceptions=BaseException, - Exception diff --git a/.readthedocs.yml b/.readthedocs.yml index 266a2cdeb2..355e3485e7 100644 --- a/.readthedocs.yml +++ b/.readthedocs.yml @@ -9,13 +9,13 @@ version: 2 build: os: ubuntu-22.04 tools: - python: "3.11" - -# Optionally set the version of Python and requirements required to build your docs -python: - install: - - requirements: docs/requirements.txt - - requirements: requirements.txt + python: "3.12" + commands: + - asdf plugin add uv + - asdf install uv latest + - asdf global uv latest + - uv sync --group docs + - uv run -m sphinx -T -b html -d docs/_build/doctrees -D language=en docs/source $READTHEDOCS_OUTPUT/html # Build documentation in the docs/ directory with Sphinx sphinx: diff --git a/CITATION.cff b/CITATION.cff index 6b0c8f19af..0be0d75d78 100644 --- a/CITATION.cff +++ b/CITATION.cff @@ -10,11 +10,11 @@ authors: version: 1.4 doi: 10.5281/zenodo.6334862 license: "MPL-2.0" -url: "https://www.coqui.ai" -repository-code: "https://github.com/coqui-ai/TTS" +url: "https://github.com/idiap/coqui-ai-TTS" +repository-code: "https://github.com/idiap/coqui-ai-TTS" keywords: - machine learning - deep learning - artificial intelligence - text to speech - - TTS \ No newline at end of file + - TTS diff --git a/CODE_OF_CONDUCT.md b/CODE_OF_CONDUCT.md index b80639d63c..9c83ebcf12 100644 --- a/CODE_OF_CONDUCT.md +++ b/CODE_OF_CONDUCT.md @@ -119,11 +119,11 @@ This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 2.0, available at [https://www.contributor-covenant.org/version/2/0/code_of_conduct.html][v2.0]. -Community Impact Guidelines were inspired by +Community Impact Guidelines were inspired by [Mozilla's code of conduct enforcement ladder][Mozilla CoC]. For answers to common questions about this code of conduct, see the FAQ at -[https://www.contributor-covenant.org/faq][FAQ]. Translations are available +[https://www.contributor-covenant.org/faq][FAQ]. Translations are available at [https://www.contributor-covenant.org/translations][translations]. [homepage]: https://www.contributor-covenant.org diff --git a/CODE_OWNERS.rst b/CODE_OWNERS.rst deleted file mode 100644 index 768b573911..0000000000 --- a/CODE_OWNERS.rst +++ /dev/null @@ -1,75 +0,0 @@ -TTS code owners / governance system -========================================== - -TTS is run under a governance system inspired (and partially copied from) by the `Mozilla module ownership system `_. The project is roughly divided into modules, and each module has its owners, which are responsible for reviewing pull requests and deciding on technical direction for their modules. Module ownership authority is given to people who have worked extensively on areas of the project. - -Module owners also have the authority of naming other module owners or appointing module peers, which are people with authority to review pull requests in that module. They can also sub-divide their module into sub-modules with their owners. - -Module owners are not tyrants. They are chartered to make decisions with input from the community and in the best interest of the community. Module owners are not required to make code changes or additions solely because the community wants them to do so. (Like anyone else, the module owners may write code because they want to, because their employers want them to, because the community wants them to, or for some other reason.) Module owners do need to pay attention to patches submitted to that module. However “pay attention” does not mean agreeing to every patch. Some patches may not make sense for the WebThings project; some may be poorly implemented. Module owners have the authority to decline a patch; this is a necessary part of the role. We ask the module owners to describe in the relevant issue their reasons for wanting changes to a patch, for declining it altogether, or for postponing review for some period. We don’t ask or expect them to rewrite patches to make them acceptable. Similarly, module owners may need to delay review of a promising patch due to an upcoming deadline. For example, a patch may be of interest, but not for the next milestone. In such a case it may make sense for the module owner to postpone review of a patch until after matters needed for a milestone have been finalized. Again, we expect this to be described in the relevant issue. And of course, it shouldn’t go on very often or for very long or escalation and review is likely. - -The work of the various module owners and peers is overseen by the global owners, which are responsible for making final decisions in case there's conflict between owners as well as set the direction for the project as a whole. - -This file describes module owners who are active on the project and which parts of the code they have expertise on (and interest in). If you're making changes to the code and are wondering who's an appropriate person to talk to, this list will tell you who to ping. - -There's overlap in the areas of expertise of each owner, and in particular when looking at which files are covered by each area, there is a lot of overlap. Don't worry about getting it exactly right when requesting review, any code owner will be happy to redirect the request to a more appropriate person. - -Global owners ----------------- - -These are people who have worked on the project extensively and are familiar with all or most parts of it. Their expertise and review guidance is trusted by other code owners to cover their own areas of expertise. In case of conflicting opinions from other owners, global owners will make a final decision. - -- Eren GÃļlge (@erogol) -- Reuben Morais (@reuben) - -Training, feeding ------------------ - -- Eren GÃļlge (@erogol) - -Model exporting ---------------- - -- Eren GÃļlge (@erogol) - -Multi-Speaker TTS ------------------ - -- Eren GÃļlge (@erogol) -- Edresson Casanova (@edresson) - -TTS ---- - -- Eren GÃļlge (@erogol) - -Vocoders --------- - -- Eren GÃļlge (@erogol) - -Speaker Encoder ---------------- - -- Eren GÃļlge (@erogol) - -Testing & CI ------------- - -- Eren GÃļlge (@erogol) -- Reuben Morais (@reuben) - -Python bindings ---------------- - -- Eren GÃļlge (@erogol) -- Reuben Morais (@reuben) - -Documentation -------------- - -- Eren GÃļlge (@erogol) - -Third party bindings --------------------- - -Owned by the author. diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index ae0ce46048..2b3a973763 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -2,7 +2,7 @@ Welcome to the 🐸TTS! -This repository is governed by [the Contributor Covenant Code of Conduct](https://github.com/coqui-ai/TTS/blob/main/CODE_OF_CONDUCT.md). +This repository is governed by [the Contributor Covenant Code of Conduct](https://github.com/idiap/coqui-ai-TTS/blob/main/CODE_OF_CONDUCT.md). ## Where to start. We welcome everyone who likes to contribute to 🐸TTS. @@ -11,30 +11,25 @@ You can contribute not only with code but with bug reports, comments, questions, If you like to contribute code, squash a bug but if you don't know where to start, here are some pointers. -- [Development Road Map](https://github.com/coqui-ai/TTS/issues/378) - - You can pick something out of our road map. We keep the progess of the project in this simple issue thread. It has new model proposals or developmental updates etc. - -- [Github Issues Tracker](https://github.com/coqui-ai/TTS/issues) +- [Github Issues Tracker](https://github.com/idiap/coqui-ai-TTS/issues) This is a place to find feature requests, bugs. - Issues with the ```good first issue``` tag are good place for beginners to take on. - -- ✨**PR**✨ [pages](https://github.com/coqui-ai/TTS/pulls) with the ```🚀new version``` tag. - - We list all the target improvements for the next version. You can pick one of them and start contributing. + Issues with the ```good first issue``` tag are good place for beginners to + take on. Issues tagged with `help wanted` are suited for more experienced + outside contributors. - Also feel free to suggest new features, ideas and models. We're always open for new things. -## Call for sharing language models +## Call for sharing pretrained models If possible, please consider sharing your pre-trained models in any language (if the licences allow for you to do so). We will include them in our model catalogue for public use and give the proper attribution, whether it be your name, company, website or any other source specified. This model can be shared in two ways: 1. Share the model files with us and we serve them with the next 🐸 TTS release. 2. Upload your models on GDrive and share the link. -Models are served under `.models.json` file and any model is available under TTS CLI or Server end points. +Models are served under `.models.json` file and any model is available under TTS +CLI and Python API end points. Either way you choose, please make sure you send the models [here](https://github.com/coqui-ai/TTS/discussions/930). @@ -44,29 +39,37 @@ If you have a new feature, a model to implement, or a bug to squash, go ahead an Please use the following steps to send a ✨**PR**✨. Let us know if you encounter a problem along the way. -The following steps are tested on an Ubuntu system. +The following steps are tested on an Ubuntu system and require +[uv](https://docs.astral.sh/uv/) for virtual environment management. Choose your +preferred [installation +method](https://docs.astral.sh/uv/getting-started/installation/), e.g. the +standalone installer: + +```bash +curl -LsSf https://astral.sh/uv/install.sh | sh +``` -1. Fork 🐸TTS[https://github.com/coqui-ai/TTS] by clicking the fork button at the top right corner of the project page. +1. Fork 🐸TTS[https://github.com/idiap/coqui-ai-TTS] by clicking the fork button at the top right corner of the project page. 2. Clone 🐸TTS and add the main repo as a new remote named ```upstream```. ```bash - $ git clone git@github.com:/TTS.git - $ cd TTS - $ git remote add upstream https://github.com/coqui-ai/TTS.git + git clone git@github.com:/coqui-ai-TTS.git + cd coqui-ai-TTS + git remote add upstream https://github.com/idiap/coqui-ai-TTS.git ``` 3. Install 🐸TTS for development. ```bash - $ make system-deps # intended to be used on Ubuntu (Debian). Let us know if you have a different OS. - $ make install + make system-deps # intended to be used on Ubuntu (Debian). Let us know if you have a different OS. + make install_dev ``` 4. Create a new branch with an informative name for your goal. ```bash - $ git checkout -b an_informative_name_for_my_branch + git checkout -b an_informative_name_for_my_branch ``` 5. Implement your changes on your new branch. @@ -75,39 +78,42 @@ The following steps are tested on an Ubuntu system. 7. Add your tests to our test suite under ```tests``` folder. It is important to show that your code works, edge cases are considered, and inform others about the intended use. -8. Run the tests to see how your updates work with the rest of the project. You can repeat this step multiple times as you implement your changes to make sure you are on the right direction. +8. Run the tests to see how your updates work with the rest of the project. You + can repeat this step multiple times as you implement your changes to make + sure you are on the right direction. **NB: running all tests takes a long time, + it is better to leave this to the CI.** ```bash - $ make test # stop at the first error - $ make test_all # run all the tests, report all the errors + uv run make test # stop at the first error + uv run make test_all # run all the tests, report all the errors ``` -9. Format your code. We use ```black``` for code and ```isort``` for ```import``` formatting. +9. Format your code. We use ```black``` for code formatting. ```bash - $ make style + make style ``` -10. Run the linter and correct the issues raised. We use ```pylint``` for linting. It helps to enforce a coding standard, offers simple refactoring suggestions. +10. Run the linter and correct the issues raised. We use ```ruff``` for linting. It helps to enforce a coding standard, offers simple refactoring suggestions. ```bash - $ make lint + make lint ``` 11. When things are good, add new files and commit your changes. ```bash - $ git add my_file1.py my_file2.py ... - $ git commit + git add my_file1.py my_file2.py ... + git commit ``` It's a good practice to regularly sync your local copy of the project with the upstream code to keep up with the recent updates. ```bash - $ git fetch upstream - $ git rebase upstream/master + git fetch upstream + git rebase upstream/main # or for the development version - $ git rebase upstream/dev + git rebase upstream/dev ``` 12. Send a PR to ```dev``` branch. @@ -115,7 +121,7 @@ The following steps are tested on an Ubuntu system. Push your branch to your fork. ```bash - $ git push -u origin an_informative_name_for_my_branch + git push -u origin an_informative_name_for_my_branch ``` Then go to your fork's Github page and click on 'Pull request' to send your ✨**PR**✨. @@ -124,7 +130,8 @@ The following steps are tested on an Ubuntu system. 13. Let's discuss until it is perfect. đŸ’Ē - We might ask you for certain changes that would appear in the ✨**PR**✨'s page under 🐸TTS[https://github.com/coqui-ai/TTS/pulls]. + We might ask you for certain changes that would appear in the + [Github ✨**PR**✨'s page](https://github.com/idiap/coqui-ai-TTS/pulls). 14. Once things look perfect, We merge it to the ```dev``` branch and make it ready for the next version. @@ -132,14 +139,14 @@ The following steps are tested on an Ubuntu system. If you prefer working within a Docker container as your development environment, you can do the following: -1. Fork 🐸TTS[https://github.com/coqui-ai/TTS] by clicking the fork button at the top right corner of the project page. +1. Fork the 🐸TTS [Github repository](https://github.com/idiap/coqui-ai-TTS) by clicking the fork button at the top right corner of the page. -2. Clone 🐸TTS and add the main repo as a new remote named ```upsteam```. +2. Clone 🐸TTS and add the main repo as a new remote named ```upstream```. ```bash - $ git clone git@github.com:/TTS.git - $ cd TTS - $ git remote add upstream https://github.com/coqui-ai/TTS.git + git clone git@github.com:/coqui-ai-TTS.git + cd coqui-ai-TTS + git remote add upstream https://github.com/idiap/coqui-ai-TTS.git ``` 3. Build the Docker Image as your development environment (it installs all of the dependencies for you): diff --git a/Dockerfile b/Dockerfile index 9fb3005ef4..9ce5c63989 100644 --- a/Dockerfile +++ b/Dockerfile @@ -1,8 +1,20 @@ ARG BASE=nvidia/cuda:11.8.0-base-ubuntu22.04 FROM ${BASE} -RUN apt-get update && apt-get upgrade -y -RUN apt-get install -y --no-install-recommends gcc g++ make python3 python3-dev python3-pip python3-venv python3-wheel espeak-ng libsndfile1-dev && rm -rf /var/lib/apt/lists/* +RUN apt-get update && \ + apt-get upgrade -y +RUN apt-get install -y --no-install-recommends \ + gcc g++ make python3 python3-dev python3-pip \ + python3-venv python3-wheel espeak-ng \ + libsndfile1-dev libc-dev curl && \ + rm -rf /var/lib/apt/lists/* + +# Install Rust compiler (to build sudachipy for Mac) +RUN curl --proto '=https' --tlsv1.2 -sSf "https://sh.rustup.rs" | sh -s -- -y +ENV PATH="/root/.cargo/bin:${PATH}" + +RUN pip3 install -U pip setuptools wheel +RUN pip3 install -U "spacy[ja]<3.8" RUN pip3 install llvmlite --ignore-installed # Install Dependencies: @@ -13,7 +25,7 @@ RUN rm -rf /root/.cache/pip WORKDIR /root COPY . /root -RUN make install +RUN pip3 install -e .[all] ENTRYPOINT ["tts"] CMD ["--help"] diff --git a/LICENSE.txt b/LICENSE.txt index 14e2f777f6..a612ad9813 100644 --- a/LICENSE.txt +++ b/LICENSE.txt @@ -35,7 +35,7 @@ Mozilla Public License Version 2.0 means any form of the work other than Source Code Form. 1.7. "Larger Work" - means a work that combines Covered Software with other material, in + means a work that combines Covered Software with other material, in a separate file or files, that is not Covered Software. 1.8. "License" diff --git a/MANIFEST.in b/MANIFEST.in deleted file mode 100644 index 321d3999c1..0000000000 --- a/MANIFEST.in +++ /dev/null @@ -1,15 +0,0 @@ -include README.md -include LICENSE.txt -include requirements.*.txt -include *.cff -include requirements.txt -include TTS/VERSION -recursive-include TTS *.json -recursive-include TTS *.html -recursive-include TTS *.png -recursive-include TTS *.md -recursive-include TTS *.py -recursive-include TTS *.pyx -recursive-include images *.png -recursive-exclude tests * -prune tests* diff --git a/Makefile b/Makefile index 7446848f46..6964773fb5 100644 --- a/Makefile +++ b/Makefile @@ -1,5 +1,5 @@ .DEFAULT_GOAL := help -.PHONY: test system-deps dev-deps deps style lint install help docs +.PHONY: test system-deps style lint install install_dev help docs help: @grep -E '^[a-zA-Z_-]+:.*?## .*$$' $(MAKEFILE_LIST) | sort | awk 'BEGIN {FS = ":.*?## "}; {printf "\033[36m%-30s\033[0m %s\n", $$1, $$2}' @@ -11,68 +11,60 @@ test_all: ## run tests and don't stop on an error. ./run_bash_tests.sh test: ## run tests. - nose2 -F -v -B --with-coverage --coverage TTS tests + coverage run -m nose2 -F -v -B tests test_vocoder: ## run vocoder tests. - nose2 -F -v -B --with-coverage --coverage TTS tests.vocoder_tests + coverage run -m nose2 -F -v -B tests.vocoder_tests test_tts: ## run tts tests. - nose2 -F -v -B --with-coverage --coverage TTS tests.tts_tests + coverage run -m nose2 -F -v -B tests.tts_tests test_tts2: ## run tts tests. - nose2 -F -v -B --with-coverage --coverage TTS tests.tts_tests2 + coverage run -m nose2 -F -v -B tests.tts_tests2 test_xtts: - nose2 -F -v -B --with-coverage --coverage TTS tests.xtts_tests + coverage run -m nose2 -F -v -B tests.xtts_tests test_aux: ## run aux tests. - nose2 -F -v -B --with-coverage --coverage TTS tests.aux_tests + coverage run -m nose2 -F -v -B tests.aux_tests ./run_bash_tests.sh -test_zoo: ## run zoo tests. - nose2 -F -v -B --with-coverage --coverage TTS tests.zoo_tests +test_zoo0: ## run zoo tests. + coverage run -m nose2 -F -v -B tests.zoo_tests.test_models.test_models_offset_0_step_3 \ + tests.zoo_tests.test_models.test_voice_conversion +test_zoo1: ## run zoo tests. + coverage run -m nose2 -F -v -B tests.zoo_tests.test_models.test_models_offset_1_step_3 +test_zoo2: ## run zoo tests. + coverage run -m nose2 -F -v -B tests.zoo_tests.test_models.test_models_offset_2_step_3 inference_tests: ## run inference tests. - nose2 -F -v -B --with-coverage --coverage TTS tests.inference_tests + coverage run -m nose2 -F -v -B tests.inference_tests data_tests: ## run data tests. - nose2 -F -v -B --with-coverage --coverage TTS tests.data_tests + coverage run -m nose2 -F -v -B tests.data_tests test_text: ## run text tests. - nose2 -F -v -B --with-coverage --coverage TTS tests.text_tests + coverage run -m nose2 -F -v -B tests.text_tests test_failed: ## only run tests failed the last time. - nose2 -F -v -B --with-coverage --coverage TTS tests + coverage run -m nose2 -F -v -B tests style: ## update code style. - black ${target_dirs} - isort ${target_dirs} + uv run --only-dev black ${target_dirs} -lint: ## run pylint linter. - pylint ${target_dirs} - black ${target_dirs} --check - isort ${target_dirs} --check-only +lint: ## run linters. + uv run --only-dev ruff check ${target_dirs} + uv run --only-dev black ${target_dirs} --check system-deps: ## install linux system deps sudo apt-get install -y libsndfile1-dev -dev-deps: ## install development deps - pip install -r requirements.dev.txt +install: ## install 🐸 TTS + uv sync --all-extras -doc-deps: ## install docs dependencies - pip install -r docs/requirements.txt - -build-docs: ## build the docs - cd docs && make clean && make build - -hub-deps: ## install deps for torch hub use - pip install -r requirements.hub.txt - -deps: ## install 🐸 requirements. - pip install -r requirements.txt - -install: ## install 🐸 TTS for development. - pip install -e .[all] +install_dev: ## install 🐸 TTS for development. + uv sync --all-extras + uv run pre-commit install docs: ## build the docs - $(MAKE) -C docs clean && $(MAKE) -C docs html + uv run --group docs $(MAKE) -C docs clean && uv run --group docs $(MAKE) -C docs html diff --git a/README.md b/README.md index e3205c1bd3..9ccf8657ab 100644 --- a/README.md +++ b/README.md @@ -1,177 +1,173 @@ +# -## 🐸Coqui.ai News -- đŸ“Ŗ ⓍTTSv2 is here with 16 languages and better performance across the board. -- đŸ“Ŗ ⓍTTS fine-tuning code is out. Check the [example recipes](https://github.com/coqui-ai/TTS/tree/dev/recipes/ljspeech). -- đŸ“Ŗ ⓍTTS can now stream with <200ms latency. -- đŸ“Ŗ ⓍTTS, our production TTS model that can speak 13 languages, is released [Blog Post](https://coqui.ai/blog/tts/open_xtts), [Demo](https://huggingface.co/spaces/coqui/xtts), [Docs](https://tts.readthedocs.io/en/dev/models/xtts.html) -- đŸ“Ŗ [đŸļBark](https://github.com/suno-ai/bark) is now available for inference with unconstrained voice cloning. [Docs](https://tts.readthedocs.io/en/dev/models/bark.html) -- đŸ“Ŗ You can use [~1100 Fairseq models](https://github.com/facebookresearch/fairseq/tree/main/examples/mms) with 🐸TTS. -- đŸ“Ŗ 🐸TTS now supports đŸĸTortoise with faster inference. [Docs](https://tts.readthedocs.io/en/dev/models/tortoise.html) -
- - -## - - -**🐸TTS is a library for advanced Text-to-Speech generation.** +**🐸 Coqui TTS is a library for advanced Text-to-Speech generation.** 🚀 Pretrained models in +1100 languages. 🛠ī¸ Tools for training new models and fine-tuning existing models in any language. 📚 Utilities for dataset analysis and curation. -______________________________________________________________________ [![Discord](https://img.shields.io/discord/1037326658807533628?color=%239B59B6&label=chat%20on%20discord)](https://discord.gg/5eXr5seRrv) +[![PyPI - Python Version](https://img.shields.io/pypi/pyversions/coqui-tts)](https://pypi.org/project/coqui-tts/) [![License]()](https://opensource.org/licenses/MPL-2.0) -[![PyPI version](https://badge.fury.io/py/TTS.svg)](https://badge.fury.io/py/TTS) -[![Covenant](https://camo.githubusercontent.com/7d620efaa3eac1c5b060ece5d6aacfcc8b81a74a04d05cd0398689c01c4463bb/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f436f6e7472696275746f72253230436f76656e616e742d76322e3025323061646f707465642d6666363962342e737667)](https://github.com/coqui-ai/TTS/blob/master/CODE_OF_CONDUCT.md) -[![Downloads](https://pepy.tech/badge/tts)](https://pepy.tech/project/tts) +[![PyPI version](https://badge.fury.io/py/coqui-tts.svg)](https://pypi.org/project/coqui-tts/) +[![Downloads](https://pepy.tech/badge/coqui-tts)](https://pepy.tech/project/coqui-tts) [![DOI](https://zenodo.org/badge/265612440.svg)](https://zenodo.org/badge/latestdoi/265612440) - -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/aux_tests.yml/badge.svg) -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/data_tests.yml/badge.svg) -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/docker.yaml/badge.svg) -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/inference_tests.yml/badge.svg) -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/style_check.yml/badge.svg) -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/text_tests.yml/badge.svg) -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/tts_tests.yml/badge.svg) -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/vocoder_tests.yml/badge.svg) -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/zoo_tests0.yml/badge.svg) -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/zoo_tests1.yml/badge.svg) -![GithubActions](https://github.com/coqui-ai/TTS/actions/workflows/zoo_tests2.yml/badge.svg) -[![Docs]()](https://tts.readthedocs.io/en/latest/) +[![GithubActions](https://github.com/idiap/coqui-ai-TTS/actions/workflows/tests.yml/badge.svg)](https://github.com/idiap/coqui-ai-TTS/actions/workflows/tests.yml) +[![GithubActions](https://github.com/idiap/coqui-ai-TTS/actions/workflows/docker.yaml/badge.svg)](https://github.com/idiap/coqui-ai-TTS/actions/workflows/docker.yaml) +[![GithubActions](https://github.com/idiap/coqui-ai-TTS/actions/workflows/style_check.yml/badge.svg)](https://github.com/idiap/coqui-ai-TTS/actions/workflows/style_check.yml) +[![Docs]()](https://coqui-tts.readthedocs.io/en/latest/)
-______________________________________________________________________ +## đŸ“Ŗ News +- **Fork of the [original, unmaintained repository](https://github.com/coqui-ai/TTS). New PyPI package: [coqui-tts](https://pypi.org/project/coqui-tts)** +- 0.25.0: [OpenVoice](https://github.com/myshell-ai/OpenVoice) models now available for voice conversion. +- 0.24.2: Prebuilt wheels are now also published for Mac and Windows (in addition to Linux as before) for easier installation across platforms. +- 0.20.0: XTTSv2 is here with 17 languages and better performance across the board. XTTS can stream with <200ms latency. +- 0.19.0: XTTS fine-tuning code is out. Check the [example recipes](https://github.com/idiap/coqui-ai-TTS/tree/dev/recipes/ljspeech). +- 0.14.1: You can use [Fairseq models in ~1100 languages](https://github.com/facebookresearch/fairseq/tree/main/examples/mms) with 🐸TTS. ## đŸ’Ŧ Where to ask questions Please use our dedicated channels for questions and discussion. Help is much more valuable if it's shared publicly so that more people can benefit from it. -| Type | Platforms | -| ------------------------------- | --------------------------------------- | -| 🚨 **Bug Reports** | [GitHub Issue Tracker] | -| 🎁 **Feature Requests & Ideas** | [GitHub Issue Tracker] | -| 👩‍đŸ’ģ **Usage Questions** | [GitHub Discussions] | -| đŸ—¯ **General Discussion** | [GitHub Discussions] or [Discord] | +| Type | Platforms | +| -------------------------------------------- | ----------------------------------- | +| 🚨 **Bug Reports, Feature Requests & Ideas** | [GitHub Issue Tracker] | +| 👩‍đŸ’ģ **Usage Questions** | [GitHub Discussions] | +| đŸ—¯ **General Discussion** | [GitHub Discussions] or [Discord] | -[github issue tracker]: https://github.com/coqui-ai/tts/issues -[github discussions]: https://github.com/coqui-ai/TTS/discussions +[github issue tracker]: https://github.com/idiap/coqui-ai-TTS/issues +[github discussions]: https://github.com/idiap/coqui-ai-TTS/discussions [discord]: https://discord.gg/5eXr5seRrv [Tutorials and Examples]: https://github.com/coqui-ai/TTS/wiki/TTS-Notebooks-and-Tutorials +The [issues](https://github.com/coqui-ai/TTS/issues) and +[discussions](https://github.com/coqui-ai/TTS/discussions) in the original +repository are also still a useful source of information. + ## 🔗 Links and Resources | Type | Links | | ------------------------------- | --------------------------------------- | -| đŸ’ŧ **Documentation** | [ReadTheDocs](https://tts.readthedocs.io/en/latest/) -| 💾 **Installation** | [TTS/README.md](https://github.com/coqui-ai/TTS/tree/dev#installation)| -| 👩‍đŸ’ģ **Contributing** | [CONTRIBUTING.md](https://github.com/coqui-ai/TTS/blob/main/CONTRIBUTING.md)| -| 📌 **Road Map** | [Main Development Plans](https://github.com/coqui-ai/TTS/issues/378) -| 🚀 **Released Models** | [TTS Releases](https://github.com/coqui-ai/TTS/releases) and [Experimental Models](https://github.com/coqui-ai/TTS/wiki/Experimental-Released-Models)| -| 📰 **Papers** | [TTS Papers](https://github.com/erogol/TTS-papers)| - - -## đŸĨ‡ TTS Performance -

- -Underlined "TTS*" and "Judy*" are **internal** 🐸TTS models that are not released open-source. They are here to show the potential. Models prefixed with a dot (.Jofish .Abe and .Janice) are real human voices. +| đŸ’ŧ **Documentation** | [ReadTheDocs](https://coqui-tts.readthedocs.io/en/latest/) +| 💾 **Installation** | [TTS/README.md](https://github.com/idiap/coqui-ai-TTS/tree/dev#installation)| +| 👩‍đŸ’ģ **Contributing** | [CONTRIBUTING.md](https://github.com/idiap/coqui-ai-TTS/blob/main/CONTRIBUTING.md)| +| 🚀 **Released Models** | [Standard models](https://github.com/idiap/coqui-ai-TTS/blob/dev/TTS/.models.json) and [Fairseq models in ~1100 languages](https://github.com/idiap/coqui-ai-TTS#example-text-to-speech-using-fairseq-models-in-1100-languages-)| ## Features -- High-performance Deep Learning models for Text2Speech tasks. - - Text2Spec models (Tacotron, Tacotron2, Glow-TTS, SpeedySpeech). - - Speaker Encoder to compute speaker embeddings efficiently. - - Vocoder models (MelGAN, Multiband-MelGAN, GAN-TTS, ParallelWaveGAN, WaveGrad, WaveRNN) -- Fast and efficient model training. -- Detailed training logs on the terminal and Tensorboard. -- Support for Multi-speaker TTS. -- Efficient, flexible, lightweight but feature complete `Trainer API`. +- High-performance text-to-speech and voice conversion models, see list below. +- Fast and efficient model training with detailed training logs on the terminal and Tensorboard. +- Support for multi-speaker and multilingual TTS. - Released and ready-to-use models. -- Tools to curate Text2Speech datasets under```dataset_analysis```. -- Utilities to use and test your models. +- Tools to curate TTS datasets under ```dataset_analysis/```. +- Command line and Python APIs to use and test your models. - Modular (but not too much) code base enabling easy implementation of new ideas. ## Model Implementations ### Spectrogram models -- Tacotron: [paper](https://arxiv.org/abs/1703.10135) -- Tacotron2: [paper](https://arxiv.org/abs/1712.05884) -- Glow-TTS: [paper](https://arxiv.org/abs/2005.11129) -- Speedy-Speech: [paper](https://arxiv.org/abs/2008.03802) -- Align-TTS: [paper](https://arxiv.org/abs/2003.01950) -- FastPitch: [paper](https://arxiv.org/pdf/2006.06873.pdf) -- FastSpeech: [paper](https://arxiv.org/abs/1905.09263) -- FastSpeech2: [paper](https://arxiv.org/abs/2006.04558) -- SC-GlowTTS: [paper](https://arxiv.org/abs/2104.05557) -- Capacitron: [paper](https://arxiv.org/abs/1906.03402) -- OverFlow: [paper](https://arxiv.org/abs/2211.06892) -- Neural HMM TTS: [paper](https://arxiv.org/abs/2108.13320) -- Delightful TTS: [paper](https://arxiv.org/abs/2110.12612) +- [Tacotron](https://arxiv.org/abs/1703.10135), [Tacotron2](https://arxiv.org/abs/1712.05884) +- [Glow-TTS](https://arxiv.org/abs/2005.11129), [SC-GlowTTS](https://arxiv.org/abs/2104.05557) +- [Speedy-Speech](https://arxiv.org/abs/2008.03802) +- [Align-TTS](https://arxiv.org/abs/2003.01950) +- [FastPitch](https://arxiv.org/pdf/2006.06873.pdf) +- [FastSpeech](https://arxiv.org/abs/1905.09263), [FastSpeech2](https://arxiv.org/abs/2006.04558) +- [Capacitron](https://arxiv.org/abs/1906.03402) +- [OverFlow](https://arxiv.org/abs/2211.06892) +- [Neural HMM TTS](https://arxiv.org/abs/2108.13320) +- [Delightful TTS](https://arxiv.org/abs/2110.12612) ### End-to-End Models -- ⓍTTS: [blog](https://coqui.ai/blog/tts/open_xtts) -- VITS: [paper](https://arxiv.org/pdf/2106.06103) -- 🐸 YourTTS: [paper](https://arxiv.org/abs/2112.02418) -- đŸĸ Tortoise: [orig. repo](https://github.com/neonbjb/tortoise-tts) -- đŸļ Bark: [orig. repo](https://github.com/suno-ai/bark) - -### Attention Methods -- Guided Attention: [paper](https://arxiv.org/abs/1710.08969) -- Forward Backward Decoding: [paper](https://arxiv.org/abs/1907.09006) -- Graves Attention: [paper](https://arxiv.org/abs/1910.10288) -- Double Decoder Consistency: [blog](https://erogol.com/solving-attention-problems-of-tts-models-with-double-decoder-consistency/) -- Dynamic Convolutional Attention: [paper](https://arxiv.org/pdf/1910.10288.pdf) -- Alignment Network: [paper](https://arxiv.org/abs/2108.10447) - -### Speaker Encoder -- GE2E: [paper](https://arxiv.org/abs/1710.10467) -- Angular Loss: [paper](https://arxiv.org/pdf/2003.11982.pdf) +- [XTTS](https://arxiv.org/abs/2406.04904) +- [VITS](https://arxiv.org/pdf/2106.06103) +- 🐸[YourTTS](https://arxiv.org/abs/2112.02418) +- đŸĸ[Tortoise](https://github.com/neonbjb/tortoise-tts) +- đŸļ[Bark](https://github.com/suno-ai/bark) ### Vocoders -- MelGAN: [paper](https://arxiv.org/abs/1910.06711) -- MultiBandMelGAN: [paper](https://arxiv.org/abs/2005.05106) -- ParallelWaveGAN: [paper](https://arxiv.org/abs/1910.11480) -- GAN-TTS discriminators: [paper](https://arxiv.org/abs/1909.11646) -- WaveRNN: [origin](https://github.com/fatchord/WaveRNN/) -- WaveGrad: [paper](https://arxiv.org/abs/2009.00713) -- HiFiGAN: [paper](https://arxiv.org/abs/2010.05646) -- UnivNet: [paper](https://arxiv.org/abs/2106.07889) +- [MelGAN](https://arxiv.org/abs/1910.06711) +- [MultiBandMelGAN](https://arxiv.org/abs/2005.05106) +- [ParallelWaveGAN](https://arxiv.org/abs/1910.11480) +- [GAN-TTS discriminators](https://arxiv.org/abs/1909.11646) +- [WaveRNN](https://github.com/fatchord/WaveRNN/) +- [WaveGrad](https://arxiv.org/abs/2009.00713) +- [HiFiGAN](https://arxiv.org/abs/2010.05646) +- [UnivNet](https://arxiv.org/abs/2106.07889) ### Voice Conversion -- FreeVC: [paper](https://arxiv.org/abs/2210.15418) +- [FreeVC](https://arxiv.org/abs/2210.15418) +- [OpenVoice](https://arxiv.org/abs/2312.01479) + +### Others +- Attention methods: [Guided Attention](https://arxiv.org/abs/1710.08969), + [Forward Backward Decoding](https://arxiv.org/abs/1907.09006), + [Graves Attention](https://arxiv.org/abs/1910.10288), + [Double Decoder Consistency](https://erogol.com/solving-attention-problems-of-tts-models-with-double-decoder-consistency/), + [Dynamic Convolutional Attention](https://arxiv.org/pdf/1910.10288.pdf), + [Alignment Network](https://arxiv.org/abs/2108.10447) +- Speaker encoders: [GE2E](https://arxiv.org/abs/1710.10467), + [Angular Loss](https://arxiv.org/pdf/2003.11982.pdf) You can also help us implement more models. + ## Installation -🐸TTS is tested on Ubuntu 18.04 with **python >= 3.9, < 3.12.**. -If you are only interested in [synthesizing speech](https://tts.readthedocs.io/en/latest/inference.html) with the released 🐸TTS models, installing from PyPI is the easiest option. +🐸TTS is tested on Ubuntu 24.04 with **python >= 3.9, < 3.13**, but should also +work on Mac and Windows. + +If you are only interested in [synthesizing speech](https://coqui-tts.readthedocs.io/en/latest/inference.html) with the pretrained 🐸TTS models, installing from PyPI is the easiest option. ```bash -pip install TTS +pip install coqui-tts ``` If you plan to code or train models, clone 🐸TTS and install it locally. ```bash -git clone https://github.com/coqui-ai/TTS -pip install -e .[all,dev,notebooks] # Select the relevant extras +git clone https://github.com/idiap/coqui-ai-TTS +cd coqui-ai-TTS +pip install -e . ``` -If you are on Ubuntu (Debian), you can also run following commands for installation. +### Optional dependencies + +The following extras allow the installation of optional dependencies: + +| Name | Description | +|------|-------------| +| `all` | All optional dependencies | +| `notebooks` | Dependencies only used in notebooks | +| `server` | Dependencies to run the TTS server | +| `bn` | Bangla G2P | +| `ja` | Japanese G2P | +| `ko` | Korean G2P | +| `zh` | Chinese G2P | +| `languages` | All language-specific dependencies | + +You can install extras with one of the following commands: ```bash -$ make system-deps # intended to be used on Ubuntu (Debian). Let us know if you have a different OS. -$ make install +pip install coqui-tts[server,ja] +pip install -e .[server,ja] ``` -If you are on Windows, 👑@GuyPaddock wrote installation instructions [here](https://stackoverflow.com/questions/66726331/how-can-i-run-mozilla-tts-coqui-tts-training-with-cuda-on-a-windows-system). +### Platforms +If you are on Ubuntu (Debian), you can also run the following commands for installation. + +```bash +make system-deps +make install +``` + + ## Docker Image -You can also try TTS without install with the docker image. -Simply run the following command and you will be able to run TTS without installing it. +You can also try out Coqui TTS without installation with the docker image. +Simply run the following command and you will be able to run TTS: ```bash docker run --rm -it -p 5002:5002 --entrypoint /bin/bash ghcr.io/coqui-ai/tts-cpu @@ -180,14 +176,15 @@ python3 TTS/server/server.py --model_name tts_models/en/vctk/vits # To start a s ``` You can then enjoy the TTS server [here](http://[::1]:5002/) -More details about the docker images (like GPU support) can be found [here](https://tts.readthedocs.io/en/latest/docker_images.html) +More details about the docker images (like GPU support) can be found +[here](https://coqui-tts.readthedocs.io/en/latest/docker_images.html) ## Synthesizing speech by 🐸TTS - + ### 🐍 Python API -#### Running a multi-speaker and multi-lingual model +#### Multi-speaker and multi-lingual model ```python import torch @@ -199,44 +196,63 @@ device = "cuda" if torch.cuda.is_available() else "cpu" # List available 🐸TTS models print(TTS().list_models()) -# Init TTS +# Initialize TTS tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device) +# List speakers +print(tts.speakers) + # Run TTS -# ❗ Since this model is multi-lingual voice cloning model, we must set the target speaker_wav and language -# Text to speech list of amplitude values as output -wav = tts.tts(text="Hello world!", speaker_wav="my/cloning/audio.wav", language="en") -# Text to speech to a file -tts.tts_to_file(text="Hello world!", speaker_wav="my/cloning/audio.wav", language="en", file_path="output.wav") +# ❗ XTTS supports both, but many models allow only one of the `speaker` and +# `speaker_wav` arguments + +# TTS with list of amplitude values as output, clone the voice from `speaker_wav` +wav = tts.tts( + text="Hello world!", + speaker_wav="my/cloning/audio.wav", + language="en" +) + +# TTS to a file, use a preset speaker +tts.tts_to_file( + text="Hello world!", + speaker="Craig Gutsy", + language="en", + file_path="output.wav" +) ``` -#### Running a single speaker model +#### Single speaker model ```python -# Init TTS with the target model name -tts = TTS(model_name="tts_models/de/thorsten/tacotron2-DDC", progress_bar=False).to(device) +# Initialize TTS with the target model name +tts = TTS("tts_models/de/thorsten/tacotron2-DDC").to(device) # Run TTS tts.tts_to_file(text="Ich bin eine Testnachricht.", file_path=OUTPUT_PATH) - -# Example voice cloning with YourTTS in English, French and Portuguese -tts = TTS(model_name="tts_models/multilingual/multi-dataset/your_tts", progress_bar=False).to(device) -tts.tts_to_file("This is voice cloning.", speaker_wav="my/cloning/audio.wav", language="en", file_path="output.wav") -tts.tts_to_file("C'est le clonage de la voix.", speaker_wav="my/cloning/audio.wav", language="fr-fr", file_path="output.wav") -tts.tts_to_file("Isso Ê clonagem de voz.", speaker_wav="my/cloning/audio.wav", language="pt-br", file_path="output.wav") ``` -#### Example voice conversion +#### Voice conversion (VC) Converting the voice in `source_wav` to the voice of `target_wav` ```python -tts = TTS(model_name="voice_conversion_models/multilingual/vctk/freevc24", progress_bar=False).to("cuda") -tts.voice_conversion_to_file(source_wav="my/source.wav", target_wav="my/target.wav", file_path="output.wav") +tts = TTS("voice_conversion_models/multilingual/vctk/freevc24").to("cuda") +tts.voice_conversion_to_file( + source_wav="my/source.wav", + target_wav="my/target.wav", + file_path="output.wav" +) ``` -#### Example voice cloning together with the voice conversion model. -This way, you can clone voices by using any model in 🐸TTS. +Other available voice conversion models: +- `voice_conversion_models/multilingual/multi-dataset/openvoice_v1` +- `voice_conversion_models/multilingual/multi-dataset/openvoice_v2` + +#### Voice cloning by combining single speaker TTS model with the default VC model + +This way, you can clone voices by using any model in 🐸TTS. The FreeVC model is +used for voice conversion after synthesizing speech. ```python @@ -248,160 +264,140 @@ tts.tts_with_vc_to_file( ) ``` -#### Example text to speech using **Fairseq models in ~1100 languages** đŸ¤¯. +#### TTS using Fairseq models in ~1100 languages đŸ¤¯ For Fairseq models, use the following name format: `tts_models//fairseq/vits`. You can find the language ISO codes [here](https://dl.fbaipublicfiles.com/mms/tts/all-tts-languages.html) and learn about the Fairseq models [here](https://github.com/facebookresearch/fairseq/tree/main/examples/mms). ```python -# TTS with on the fly voice conversion +# TTS with fairseq models api = TTS("tts_models/deu/fairseq/vits") -api.tts_with_vc_to_file( +api.tts_to_file( "Wie sage ich auf Italienisch, dass ich dich liebe?", - speaker_wav="target/speaker.wav", file_path="output.wav" ) ``` -### Command-line `tts` +### Command-line interface `tts` -Synthesize speech on command line. +Synthesize speech on the command line. You can either use your trained model or choose a model from the provided list. -If you don't specify any models, then it uses LJSpeech based English model. - -#### Single Speaker Models - - List provided models: + ```sh + tts --list_models ``` - $ tts --list_models - ``` - -- Get model info (for both tts_models and vocoder_models): - - - Query by type/name: - The model_info_by_name uses the name as it from the --list_models. - ``` - $ tts --model_info_by_name "///" - ``` - For example: - ``` - $ tts --model_info_by_name tts_models/tr/common-voice/glow-tts - $ tts --model_info_by_name vocoder_models/en/ljspeech/hifigan_v2 - ``` - - Query by type/idx: - The model_query_idx uses the corresponding idx from --list_models. - - ``` - $ tts --model_info_by_idx "/" - ``` - For example: - - ``` - $ tts --model_info_by_idx tts_models/3 - ``` +- Get model information. Use the names obtained from `--list_models`. + ```sh + tts --model_info_by_name "///" + ``` + For example: + ```sh + tts --model_info_by_name tts_models/tr/common-voice/glow-tts + tts --model_info_by_name vocoder_models/en/ljspeech/hifigan_v2 + ``` - - Query info for model info by full name: - ``` - $ tts --model_info_by_name "///" - ``` +#### Single speaker models -- Run TTS with default models: +- Run TTS with the default model (`tts_models/en/ljspeech/tacotron2-DDC`): - ``` - $ tts --text "Text for TTS" --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" --out_path output/path/speech.wav ``` - Run TTS and pipe out the generated TTS wav file data: - ``` - $ tts --text "Text for TTS" --pipe_out --out_path output/path/speech.wav | aplay + ```sh + tts --text "Text for TTS" --pipe_out --out_path output/path/speech.wav | aplay ``` - Run a TTS model with its default vocoder model: - ``` - $ tts --text "Text for TTS" --model_name "///" --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" \ + --model_name "///" \ + --out_path output/path/speech.wav ``` For example: - ``` - $ tts --text "Text for TTS" --model_name "tts_models/en/ljspeech/glow-tts" --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" \ + --model_name "tts_models/en/ljspeech/glow-tts" \ + --out_path output/path/speech.wav ``` -- Run with specific TTS and vocoder models from the list: +- Run with specific TTS and vocoder models from the list. Note that not every vocoder is compatible with every TTS model. - ``` - $ tts --text "Text for TTS" --model_name "///" --vocoder_name "///" --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" \ + --model_name "///" \ + --vocoder_name "///" \ + --out_path output/path/speech.wav ``` For example: - ``` - $ tts --text "Text for TTS" --model_name "tts_models/en/ljspeech/glow-tts" --vocoder_name "vocoder_models/en/ljspeech/univnet" --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" \ + --model_name "tts_models/en/ljspeech/glow-tts" \ + --vocoder_name "vocoder_models/en/ljspeech/univnet" \ + --out_path output/path/speech.wav ``` -- Run your own TTS model (Using Griffin-Lim Vocoder): +- Run your own TTS model (using Griffin-Lim Vocoder): - ``` - $ tts --text "Text for TTS" --model_path path/to/model.pth --config_path path/to/config.json --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" \ + --model_path path/to/model.pth \ + --config_path path/to/config.json \ + --out_path output/path/speech.wav ``` - Run your own TTS and Vocoder models: - ``` - $ tts --text "Text for TTS" --model_path path/to/model.pth --config_path path/to/config.json --out_path output/path/speech.wav - --vocoder_path path/to/vocoder.pth --vocoder_config_path path/to/vocoder_config.json + ```sh + tts --text "Text for TTS" \ + --model_path path/to/model.pth \ + --config_path path/to/config.json \ + --out_path output/path/speech.wav \ + --vocoder_path path/to/vocoder.pth \ + --vocoder_config_path path/to/vocoder_config.json ``` -#### Multi-speaker Models +#### Multi-speaker models -- List the available speakers and choose a among them: +- List the available speakers and choose a `` among them: - ``` - $ tts --model_name "//" --list_speaker_idxs + ```sh + tts --model_name "//" --list_speaker_idxs ``` - Run the multi-speaker TTS model with the target speaker ID: - ``` - $ tts --text "Text for TTS." --out_path output/path/speech.wav --model_name "//" --speaker_idx + ```sh + tts --text "Text for TTS." --out_path output/path/speech.wav \ + --model_name "//" --speaker_idx ``` - Run your own multi-speaker TTS model: - ``` - $ tts --text "Text for TTS" --out_path output/path/speech.wav --model_path path/to/model.pth --config_path path/to/config.json --speakers_file_path path/to/speaker.json --speaker_idx + ```sh + tts --text "Text for TTS" --out_path output/path/speech.wav \ + --model_path path/to/model.pth --config_path path/to/config.json \ + --speakers_file_path path/to/speaker.json --speaker_idx ``` -### Voice Conversion Models +#### Voice conversion models -``` -$ tts --out_path output/path/speech.wav --model_name "//" --source_wav --target_wav +```sh +tts --out_path output/path/speech.wav --model_name "//" \ + --source_wav --target_wav ``` - -## Directory Structure -``` -|- notebooks/ (Jupyter Notebooks for model evaluation, parameter selection and data analysis.) -|- utils/ (common utilities.) -|- TTS - |- bin/ (folder for all the executables.) - |- train*.py (train your target model.) - |- ... - |- tts/ (text to speech models) - |- layers/ (model layer definitions) - |- models/ (model definitions) - |- utils/ (model specific utilities.) - |- speaker_encoder/ (Speaker Encoder models.) - |- (same) - |- vocoder/ (Vocoder models.) - |- (same) -``` diff --git a/TTS/.models.json b/TTS/.models.json index b349e7397b..36654d0555 100644 --- a/TTS/.models.json +++ b/TTS/.models.json @@ -5,11 +5,11 @@ "xtts_v2": { "description": "XTTS-v2.0.3 by Coqui with 17 languages.", "hf_url": [ - "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/model.pth", - "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/config.json", - "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/vocab.json", - "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/hash.md5", - "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/speakers_xtts.pth" + "https://huggingface.co/coqui/XTTS-v2/resolve/main/model.pth", + "https://huggingface.co/coqui/XTTS-v2/resolve/main/config.json", + "https://huggingface.co/coqui/XTTS-v2/resolve/main/vocab.json", + "https://huggingface.co/coqui/XTTS-v2/resolve/main/hash.md5", + "https://huggingface.co/coqui/XTTS-v2/resolve/main/speakers_xtts.pth" ], "model_hash": "10f92b55c512af7a8d39d650547a15a7", "default_vocoder": null, @@ -21,10 +21,10 @@ "xtts_v1.1": { "description": "XTTS-v1.1 by Coqui with 14 languages, cross-language voice cloning and reference leak fixed.", "hf_url": [ - "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v1/v1.1.2/model.pth", - "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v1/v1.1.2/config.json", - "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v1/v1.1.2/vocab.json", - "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v1/v1.1.2/hash.md5" + "https://huggingface.co/coqui/XTTS-v1/resolve/v1.1.2/model.pth", + "https://huggingface.co/coqui/XTTS-v1/resolve/v1.1.2/config.json", + "https://huggingface.co/coqui/XTTS-v1/resolve/v1.1.2/vocab.json", + "https://huggingface.co/coqui/XTTS-v1/resolve/v1.1.2/hash.md5" ], "model_hash": "7c62beaf58d39b729de287330dc254e7b515677416839b649a50e7cf74c3df59", "default_vocoder": null, @@ -35,7 +35,7 @@ }, "your_tts": { "description": "Your TTS model accompanying the paper https://arxiv.org/abs/2112.02418", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.10.1_models/tts_models--multilingual--multi-dataset--your_tts.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.10.1_models/tts_models--multilingual--multi-dataset--your_tts.zip", "default_vocoder": null, "commit": "e9a1953e", "license": "CC BY-NC-ND 4.0", @@ -44,12 +44,11 @@ "bark": { "description": "đŸļ Bark TTS model released by suno-ai. You can find the original implementation in https://github.com/suno-ai/bark.", "hf_url": [ - "https://coqui.gateway.scarf.sh/hf/bark/coarse_2.pt", - "https://coqui.gateway.scarf.sh/hf/bark/fine_2.pt", - "https://coqui.gateway.scarf.sh/hf/text_2.pt", - "https://coqui.gateway.scarf.sh/hf/bark/config.json", - "https://coqui.gateway.scarf.sh/hf/bark/hubert.pt", - "https://coqui.gateway.scarf.sh/hf/bark/tokenizer.pth" + "https://huggingface.co/erogol/bark/resolve/main/coarse_2.pt", + "https://huggingface.co/erogol/bark/resolve/main/fine_2.pt", + "https://huggingface.co/erogol/bark/resolve/main/text_2.pt", + "https://huggingface.co/erogol/bark/resolve/main/config.json", + "https://huggingface.co/erogol/bark/resolve/main/tokenizer.pth" ], "default_vocoder": null, "commit": "e9a1953e", @@ -61,7 +60,7 @@ "bg": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--bg--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--bg--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -72,7 +71,7 @@ "cs": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--cs--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--cs--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -83,7 +82,7 @@ "da": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--da--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--da--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -94,7 +93,7 @@ "et": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--et--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--et--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -105,7 +104,7 @@ "ga": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--ga--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--ga--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -117,7 +116,7 @@ "ek1": { "tacotron2": { "description": "EK1 en-rp tacotron2 by NMStoker", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--ek1--tacotron2.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--ek1--tacotron2.zip", "default_vocoder": "vocoder_models/en/ek1/wavegrad", "commit": "c802255", "license": "apache 2.0" @@ -126,7 +125,7 @@ "ljspeech": { "tacotron2-DDC": { "description": "Tacotron2 with Double Decoder Consistency.", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--ljspeech--tacotron2-DDC.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--ljspeech--tacotron2-DDC.zip", "default_vocoder": "vocoder_models/en/ljspeech/hifigan_v2", "commit": "bae2ad0f", "author": "Eren GÃļlge @erogol", @@ -135,7 +134,7 @@ }, "tacotron2-DDC_ph": { "description": "Tacotron2 with Double Decoder Consistency with phonemes.", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--ljspeech--tacotron2-DDC_ph.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--ljspeech--tacotron2-DDC_ph.zip", "default_vocoder": "vocoder_models/en/ljspeech/univnet", "commit": "3900448", "author": "Eren GÃļlge @erogol", @@ -144,7 +143,7 @@ }, "glow-tts": { "description": "", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--ljspeech--glow-tts.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--ljspeech--glow-tts.zip", "stats_file": null, "default_vocoder": "vocoder_models/en/ljspeech/multiband-melgan", "commit": "", @@ -154,7 +153,7 @@ }, "speedy-speech": { "description": "Speedy Speech model trained on LJSpeech dataset using the Alignment Network for learning the durations.", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--ljspeech--speedy-speech.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--ljspeech--speedy-speech.zip", "stats_file": null, "default_vocoder": "vocoder_models/en/ljspeech/hifigan_v2", "commit": "4581e3d", @@ -164,7 +163,7 @@ }, "tacotron2-DCA": { "description": "", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--ljspeech--tacotron2-DCA.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--ljspeech--tacotron2-DCA.zip", "default_vocoder": "vocoder_models/en/ljspeech/multiband-melgan", "commit": "", "author": "Eren GÃļlge @erogol", @@ -173,7 +172,7 @@ }, "vits": { "description": "VITS is an End2End TTS model trained on LJSpeech dataset with phonemes.", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--ljspeech--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--ljspeech--vits.zip", "default_vocoder": null, "commit": "3900448", "author": "Eren GÃļlge @erogol", @@ -181,7 +180,7 @@ "contact": "egolge@coqui.com" }, "vits--neon": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--en--ljspeech--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--en--ljspeech--vits.zip", "default_vocoder": null, "author": "@NeonGeckoCom", "license": "bsd-3-clause", @@ -190,7 +189,7 @@ }, "fast_pitch": { "description": "FastPitch model trained on LJSpeech using the Aligner Network", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--ljspeech--fast_pitch.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--ljspeech--fast_pitch.zip", "default_vocoder": "vocoder_models/en/ljspeech/hifigan_v2", "commit": "b27b3ba", "author": "Eren GÃļlge @erogol", @@ -199,7 +198,7 @@ }, "overflow": { "description": "Overflow model trained on LJSpeech", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.10.0_models/tts_models--en--ljspeech--overflow.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.10.0_models/tts_models--en--ljspeech--overflow.zip", "default_vocoder": "vocoder_models/en/ljspeech/hifigan_v2", "commit": "3b1a28f", "author": "Eren GÃļlge @erogol", @@ -208,7 +207,7 @@ }, "neural_hmm": { "description": "Neural HMM model trained on LJSpeech", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.11.0_models/tts_models--en--ljspeech--neural_hmm.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.11.0_models/tts_models--en--ljspeech--neural_hmm.zip", "default_vocoder": "vocoder_models/en/ljspeech/hifigan_v2", "commit": "3b1a28f", "author": "Shivam Metha @shivammehta25", @@ -219,7 +218,7 @@ "vctk": { "vits": { "description": "VITS End2End TTS model trained on VCTK dataset with 109 different speakers with EN accent.", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--vctk--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--vctk--vits.zip", "default_vocoder": null, "commit": "3900448", "author": "Eren @erogol", @@ -228,7 +227,7 @@ }, "fast_pitch": { "description": "FastPitch model trained on VCTK dataseset.", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--vctk--fast_pitch.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--vctk--fast_pitch.zip", "default_vocoder": null, "commit": "bdab788d", "author": "Eren @erogol", @@ -239,7 +238,7 @@ "sam": { "tacotron-DDC": { "description": "Tacotron2 with Double Decoder Consistency trained with Aceenture's Sam dataset.", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--en--sam--tacotron-DDC.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--en--sam--tacotron-DDC.zip", "default_vocoder": "vocoder_models/en/sam/hifigan_v2", "commit": "bae2ad0f", "author": "Eren GÃļlge @erogol", @@ -250,7 +249,7 @@ "blizzard2013": { "capacitron-t2-c50": { "description": "Capacitron additions to Tacotron 2 with Capacity at 50 as in https://arxiv.org/pdf/1906.03402.pdf", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.7.0_models/tts_models--en--blizzard2013--capacitron-t2-c50.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.7.0_models/tts_models--en--blizzard2013--capacitron-t2-c50.zip", "commit": "d6284e7", "default_vocoder": "vocoder_models/en/blizzard2013/hifigan_v2", "author": "Adam Froghyar @a-froghyar", @@ -259,7 +258,7 @@ }, "capacitron-t2-c150_v2": { "description": "Capacitron additions to Tacotron 2 with Capacity at 150 as in https://arxiv.org/pdf/1906.03402.pdf", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.7.1_models/tts_models--en--blizzard2013--capacitron-t2-c150_v2.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.7.1_models/tts_models--en--blizzard2013--capacitron-t2-c150_v2.zip", "commit": "a67039d", "default_vocoder": "vocoder_models/en/blizzard2013/hifigan_v2", "author": "Adam Froghyar @a-froghyar", @@ -271,15 +270,15 @@ "tortoise-v2": { "description": "Tortoise tts model https://github.com/neonbjb/tortoise-tts", "github_rls_url": [ - "https://coqui.gateway.scarf.sh/v0.14.1_models/autoregressive.pth", - "https://coqui.gateway.scarf.sh/v0.14.1_models/clvp2.pth", - "https://coqui.gateway.scarf.sh/v0.14.1_models/cvvp.pth", - "https://coqui.gateway.scarf.sh/v0.14.1_models/diffusion_decoder.pth", - "https://coqui.gateway.scarf.sh/v0.14.1_models/rlg_auto.pth", - "https://coqui.gateway.scarf.sh/v0.14.1_models/rlg_diffuser.pth", - "https://coqui.gateway.scarf.sh/v0.14.1_models/vocoder.pth", - "https://coqui.gateway.scarf.sh/v0.14.1_models/mel_norms.pth", - "https://coqui.gateway.scarf.sh/v0.14.1_models/config.json" + "https://github.com/coqui-ai/TTS/releases/download/v0.14.1_models/autoregressive.pth", + "https://github.com/coqui-ai/TTS/releases/download/v0.14.1_models/clvp2.pth", + "https://github.com/coqui-ai/TTS/releases/download/v0.14.1_models/cvvp.pth", + "https://github.com/coqui-ai/TTS/releases/download/v0.14.1_models/diffusion_decoder.pth", + "https://github.com/coqui-ai/TTS/releases/download/v0.14.1_models/rlg_auto.pth", + "https://github.com/coqui-ai/TTS/releases/download/v0.14.1_models/rlg_diffuser.pth", + "https://github.com/coqui-ai/TTS/releases/download/v0.14.1_models/vocoder.pth", + "https://github.com/coqui-ai/TTS/releases/download/v0.14.1_models/mel_norms.pth", + "https://github.com/coqui-ai/TTS/releases/download/v0.14.1_models/config.json" ], "commit": "c1875f6", "default_vocoder": null, @@ -290,7 +289,7 @@ "jenny": { "jenny": { "description": "VITS model trained with Jenny(Dioco) dataset. Named as Jenny as demanded by the license. Original URL for the model https://www.kaggle.com/datasets/noml4u/tts-models--en--jenny-dioco--vits", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.14.0_models/tts_models--en--jenny--jenny.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.14.0_models/tts_models--en--jenny--jenny.zip", "default_vocoder": null, "commit": "ba40a1c", "license": "custom - see https://github.com/dioco-group/jenny-tts-dataset#important", @@ -301,7 +300,7 @@ "es": { "mai": { "tacotron2-DDC": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--es--mai--tacotron2-DDC.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--es--mai--tacotron2-DDC.zip", "default_vocoder": "vocoder_models/universal/libri-tts/fullband-melgan", "commit": "", "author": "Eren GÃļlge @erogol", @@ -311,7 +310,7 @@ }, "css10": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--es--css10--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--es--css10--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -322,7 +321,7 @@ "fr": { "mai": { "tacotron2-DDC": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--fr--mai--tacotron2-DDC.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--fr--mai--tacotron2-DDC.zip", "default_vocoder": "vocoder_models/universal/libri-tts/fullband-melgan", "commit": null, "author": "Eren GÃļlge @erogol", @@ -332,7 +331,7 @@ }, "css10": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--fr--css10--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--fr--css10--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -343,7 +342,7 @@ "uk": { "mai": { "glow-tts": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--uk--mai--glow-tts.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--uk--mai--glow-tts.zip", "author": "@robinhad", "commit": "bdab788d", "license": "MIT", @@ -351,7 +350,7 @@ "default_vocoder": "vocoder_models/uk/mai/multiband-melgan" }, "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--uk--mai--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--uk--mai--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -362,7 +361,7 @@ "zh-CN": { "baker": { "tacotron2-DDC-GST": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--zh-CN--baker--tacotron2-DDC-GST.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--zh-CN--baker--tacotron2-DDC-GST.zip", "commit": "unknown", "author": "@kirianguiller", "license": "apache 2.0", @@ -373,7 +372,7 @@ "nl": { "mai": { "tacotron2-DDC": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--nl--mai--tacotron2-DDC.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--nl--mai--tacotron2-DDC.zip", "author": "@r-dh", "license": "apache 2.0", "default_vocoder": "vocoder_models/nl/mai/parallel-wavegan", @@ -383,7 +382,7 @@ }, "css10": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--nl--css10--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--nl--css10--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -394,21 +393,21 @@ "de": { "thorsten": { "tacotron2-DCA": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--de--thorsten--tacotron2-DCA.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--de--thorsten--tacotron2-DCA.zip", "default_vocoder": "vocoder_models/de/thorsten/fullband-melgan", "author": "@thorstenMueller", "license": "apache 2.0", "commit": "unknown" }, "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.7.0_models/tts_models--de--thorsten--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.7.0_models/tts_models--de--thorsten--vits.zip", "default_vocoder": null, "author": "@thorstenMueller", "license": "apache 2.0", "commit": "unknown" }, "tacotron2-DDC": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--de--thorsten--tacotron2-DDC.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--de--thorsten--tacotron2-DDC.zip", "default_vocoder": "vocoder_models/de/thorsten/hifigan_v1", "description": "Thorsten-Dec2021-22k-DDC", "author": "@thorstenMueller", @@ -418,7 +417,7 @@ }, "css10": { "vits-neon": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--de--css10--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--de--css10--vits.zip", "default_vocoder": null, "author": "@NeonGeckoCom", "license": "bsd-3-clause", @@ -429,7 +428,7 @@ "ja": { "kokoro": { "tacotron2-DDC": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--ja--kokoro--tacotron2-DDC.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--ja--kokoro--tacotron2-DDC.zip", "default_vocoder": "vocoder_models/ja/kokoro/hifigan_v1", "description": "Tacotron2 with Double Decoder Consistency trained with Kokoro Speech Dataset.", "author": "@kaiidams", @@ -441,7 +440,7 @@ "tr": { "common-voice": { "glow-tts": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--tr--common-voice--glow-tts.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--tr--common-voice--glow-tts.zip", "default_vocoder": "vocoder_models/tr/common-voice/hifigan", "license": "MIT", "description": "Turkish GlowTTS model using an unknown speaker from the Common-Voice dataset.", @@ -453,7 +452,7 @@ "it": { "mai_female": { "glow-tts": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--it--mai_female--glow-tts.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--it--mai_female--glow-tts.zip", "default_vocoder": null, "description": "GlowTTS model as explained on https://github.com/coqui-ai/TTS/issues/1148.", "author": "@nicolalandro", @@ -461,7 +460,7 @@ "commit": null }, "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--it--mai_female--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--it--mai_female--vits.zip", "default_vocoder": null, "description": "GlowTTS model as explained on https://github.com/coqui-ai/TTS/issues/1148.", "author": "@nicolalandro", @@ -471,7 +470,7 @@ }, "mai_male": { "glow-tts": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--it--mai_male--glow-tts.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--it--mai_male--glow-tts.zip", "default_vocoder": null, "description": "GlowTTS model as explained on https://github.com/coqui-ai/TTS/issues/1148.", "author": "@nicolalandro", @@ -479,7 +478,7 @@ "commit": null }, "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--it--mai_male--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/tts_models--it--mai_male--vits.zip", "default_vocoder": null, "description": "GlowTTS model as explained on https://github.com/coqui-ai/TTS/issues/1148.", "author": "@nicolalandro", @@ -491,7 +490,7 @@ "ewe": { "openbible": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.2_models/tts_models--ewe--openbible--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.2_models/tts_models--ewe--openbible--vits.zip", "default_vocoder": null, "license": "CC-BY-SA 4.0", "description": "Original work (audio and text) by Biblica available for free at www.biblica.com and open.bible.", @@ -503,7 +502,7 @@ "hau": { "openbible": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.2_models/tts_models--hau--openbible--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.2_models/tts_models--hau--openbible--vits.zip", "default_vocoder": null, "license": "CC-BY-SA 4.0", "description": "Original work (audio and text) by Biblica available for free at www.biblica.com and open.bible.", @@ -515,7 +514,7 @@ "lin": { "openbible": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.2_models/tts_models--lin--openbible--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.2_models/tts_models--lin--openbible--vits.zip", "default_vocoder": null, "license": "CC-BY-SA 4.0", "description": "Original work (audio and text) by Biblica available for free at www.biblica.com and open.bible.", @@ -527,7 +526,7 @@ "tw_akuapem": { "openbible": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.2_models/tts_models--tw_akuapem--openbible--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.2_models/tts_models--tw_akuapem--openbible--vits.zip", "default_vocoder": null, "license": "CC-BY-SA 4.0", "description": "Original work (audio and text) by Biblica available for free at www.biblica.com and open.bible.", @@ -539,7 +538,7 @@ "tw_asante": { "openbible": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.2_models/tts_models--tw_asante--openbible--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.2_models/tts_models--tw_asante--openbible--vits.zip", "default_vocoder": null, "license": "CC-BY-SA 4.0", "description": "Original work (audio and text) by Biblica available for free at www.biblica.com and open.bible.", @@ -551,7 +550,7 @@ "yor": { "openbible": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.2_models/tts_models--yor--openbible--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.2_models/tts_models--yor--openbible--vits.zip", "default_vocoder": null, "license": "CC-BY-SA 4.0", "description": "Original work (audio and text) by Biblica available for free at www.biblica.com and open.bible.", @@ -563,7 +562,7 @@ "hu": { "css10": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--hu--css10--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--hu--css10--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -574,7 +573,7 @@ "el": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--el--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--el--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -585,7 +584,7 @@ "fi": { "css10": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--fi--css10--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--fi--css10--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -596,7 +595,7 @@ "hr": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--hr--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--hr--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -607,7 +606,7 @@ "lt": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--lt--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--lt--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -618,7 +617,7 @@ "lv": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--lv--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--lv--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -629,7 +628,7 @@ "mt": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--mt--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--mt--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -640,7 +639,7 @@ "pl": { "mai_female": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--pl--mai_female--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--pl--mai_female--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -651,7 +650,7 @@ "pt": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--pt--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--pt--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -662,7 +661,7 @@ "ro": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--ro--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--ro--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -673,7 +672,7 @@ "sk": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--sk--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--sk--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -684,7 +683,7 @@ "sl": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--sl--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--sl--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -695,7 +694,7 @@ "sv": { "cv": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/tts_models--sv--cv--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--sv--cv--vits.zip", "default_vocoder": null, "commit": null, "author": "@NeonGeckoCom", @@ -706,7 +705,7 @@ "ca": { "custom": { "vits": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.10.1_models/tts_models--ca--custom--vits.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.10.1_models/tts_models--ca--custom--vits.zip", "default_vocoder": null, "commit": null, "description": " It is trained from zero with 101460 utterances consisting of 257 speakers, approx 138 hours of speech. We used three datasets;\nFestcat and Google Catalan TTS (both TTS datasets) and also a part of Common Voice 8. It is trained with TTS v0.8.0.\nhttps://github.com/coqui-ai/TTS/discussions/930#discussioncomment-4466345", @@ -718,7 +717,7 @@ "fa": { "custom": { "glow-tts": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.10.1_models/tts_models--fa--custom--glow-tts.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.10.1_models/tts_models--fa--custom--glow-tts.zip", "default_vocoder": null, "commit": null, "description": "persian-tts-female-glow_tts model for text to speech purposes. Single-speaker female voice Trained on persian-tts-dataset-famale. \nThis model has no compatible vocoder thus the output quality is not very good. \nDataset: https://www.kaggle.com/datasets/magnoliasis/persian-tts-dataset-famale.", @@ -730,7 +729,7 @@ "bn": { "custom": { "vits-male": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.13.3_models/tts_models--bn--custom--vits_male.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.13.3_models/tts_models--bn--custom--vits_male.zip", "default_vocoder": null, "commit": null, "description": "Single speaker Bangla male model. For more information -> https://github.com/mobassir94/comprehensive-bangla-tts", @@ -738,7 +737,7 @@ "license": "Apache 2.0" }, "vits-female": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.13.3_models/tts_models--bn--custom--vits_female.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.13.3_models/tts_models--bn--custom--vits_female.zip", "default_vocoder": null, "commit": null, "description": "Single speaker Bangla female model. For more information -> https://github.com/mobassir94/comprehensive-bangla-tts", @@ -751,7 +750,7 @@ "common-voice": { "glow-tts":{ "description": "Belarusian GlowTTS model created by @alex73 (Github).", - "github_rls_url":"https://coqui.gateway.scarf.sh/v0.16.6/tts_models--be--common-voice--glow-tts.zip", + "github_rls_url":"https://github.com/coqui-ai/TTS/releases/download/v0.16.6/tts_models--be--common-voice--glow-tts.zip", "default_vocoder": "vocoder_models/be/common-voice/hifigan", "commit": "c0aabb85", "license": "CC-BY-SA 4.0", @@ -764,14 +763,14 @@ "universal": { "libri-tts": { "wavegrad": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--universal--libri-tts--wavegrad.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--universal--libri-tts--wavegrad.zip", "commit": "ea976b0", "author": "Eren GÃļlge @erogol", "license": "MPL", "contact": "egolge@coqui.com" }, "fullband-melgan": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--universal--libri-tts--fullband-melgan.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--universal--libri-tts--fullband-melgan.zip", "commit": "4132240", "author": "Eren GÃļlge @erogol", "license": "MPL", @@ -783,14 +782,14 @@ "ek1": { "wavegrad": { "description": "EK1 en-rp wavegrad by NMStoker", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--en--ek1--wavegrad.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--en--ek1--wavegrad.zip", "commit": "c802255", "license": "apache 2.0" } }, "ljspeech": { "multiband-melgan": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--en--ljspeech--multiband-melgan.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--en--ljspeech--multiband-melgan.zip", "commit": "ea976b0", "author": "Eren GÃļlge @erogol", "license": "MPL", @@ -798,7 +797,7 @@ }, "hifigan_v2": { "description": "HiFiGAN_v2 LJSpeech vocoder from https://arxiv.org/abs/2010.05646.", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--en--ljspeech--hifigan_v2.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--en--ljspeech--hifigan_v2.zip", "commit": "bae2ad0f", "author": "@erogol", "license": "apache 2.0", @@ -806,7 +805,7 @@ }, "univnet": { "description": "UnivNet model finetuned on TacotronDDC_ph spectrograms for better compatibility.", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--en--ljspeech--univnet_v2.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--en--ljspeech--univnet_v2.zip", "commit": "4581e3d", "author": "Eren @erogol", "license": "apache 2.0", @@ -816,7 +815,7 @@ "blizzard2013": { "hifigan_v2": { "description": "HiFiGAN_v2 LJSpeech vocoder from https://arxiv.org/abs/2010.05646.", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.7.0_models/vocoder_models--en--blizzard2013--hifigan_v2.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.7.0_models/vocoder_models--en--blizzard2013--hifigan_v2.zip", "commit": "d6284e7", "author": "Adam Froghyar @a-froghyar", "license": "apache 2.0", @@ -826,7 +825,7 @@ "vctk": { "hifigan_v2": { "description": "Finetuned and intended to be used with tts_models/en/vctk/sc-glow-tts", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--en--vctk--hifigan_v2.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--en--vctk--hifigan_v2.zip", "commit": "2f07160", "author": "Edresson Casanova", "license": "apache 2.0", @@ -836,7 +835,7 @@ "sam": { "hifigan_v2": { "description": "Finetuned and intended to be used with tts_models/en/sam/tacotron_DDC", - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--en--sam--hifigan_v2.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--en--sam--hifigan_v2.zip", "commit": "2f07160", "author": "Eren GÃļlge @erogol", "license": "apache 2.0", @@ -847,7 +846,7 @@ "nl": { "mai": { "parallel-wavegan": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--nl--mai--parallel-wavegan.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--nl--mai--parallel-wavegan.zip", "author": "@r-dh", "license": "apache 2.0", "commit": "unknown" @@ -857,19 +856,19 @@ "de": { "thorsten": { "wavegrad": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--de--thorsten--wavegrad.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--de--thorsten--wavegrad.zip", "author": "@thorstenMueller", "license": "apache 2.0", "commit": "unknown" }, "fullband-melgan": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--de--thorsten--fullband-melgan.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--de--thorsten--fullband-melgan.zip", "author": "@thorstenMueller", "license": "apache 2.0", "commit": "unknown" }, "hifigan_v1": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.8.0_models/vocoder_models--de--thorsten--hifigan_v1.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/vocoder_models--de--thorsten--hifigan_v1.zip", "description": "HifiGAN vocoder model for Thorsten Neutral Dec2021 22k Samplerate Tacotron2 DDC model", "author": "@thorstenMueller", "license": "apache 2.0", @@ -880,7 +879,7 @@ "ja": { "kokoro": { "hifigan_v1": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--ja--kokoro--hifigan_v1.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--ja--kokoro--hifigan_v1.zip", "description": "HifiGAN model trained for kokoro dataset by @kaiidams", "author": "@kaiidams", "license": "apache 2.0", @@ -891,7 +890,7 @@ "uk": { "mai": { "multiband-melgan": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--uk--mai--multiband-melgan.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--uk--mai--multiband-melgan.zip", "author": "@robinhad", "commit": "bdab788d", "license": "MIT", @@ -902,7 +901,7 @@ "tr": { "common-voice": { "hifigan": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/vocoder_models--tr--common-voice--hifigan.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.6.1_models/vocoder_models--tr--common-voice--hifigan.zip", "description": "HifiGAN model using an unknown speaker from the Common-Voice dataset.", "author": "Fatih Akademi", "license": "MIT", @@ -913,7 +912,7 @@ "be": { "common-voice": { "hifigan": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.16.6/vocoder_models--be--common-voice--hifigan.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.16.6/vocoder_models--be--common-voice--hifigan.zip", "description": "Belarusian HiFiGAN model created by @alex73 (Github).", "author": "@alex73", "license": "CC-BY-SA 4.0", @@ -926,12 +925,34 @@ "multilingual": { "vctk": { "freevc24": { - "github_rls_url": "https://coqui.gateway.scarf.sh/v0.13.0_models/voice_conversion_models--multilingual--vctk--freevc24.zip", + "github_rls_url": "https://github.com/coqui-ai/TTS/releases/download/v0.13.0_models/voice_conversion_models--multilingual--vctk--freevc24.zip", "description": "FreeVC model trained on VCTK dataset from https://github.com/OlaWod/FreeVC", "author": "Jing-Yi Li @OlaWod", "license": "MIT", "commit": null } + }, + "multi-dataset": { + "openvoice_v1": { + "hf_url": [ + "https://huggingface.co/myshell-ai/OpenVoice/resolve/main/checkpoints/converter/config.json", + "https://huggingface.co/myshell-ai/OpenVoice/resolve/main/checkpoints/converter/checkpoint.pth" + ], + "description": "OpenVoice VC model from https://huggingface.co/myshell-ai/OpenVoiceV2", + "author": "MyShell.ai", + "license": "MIT", + "commit": null + }, + "openvoice_v2": { + "hf_url": [ + "https://huggingface.co/myshell-ai/OpenVoiceV2/resolve/main/converter/config.json", + "https://huggingface.co/myshell-ai/OpenVoiceV2/resolve/main/converter/checkpoint.pth" + ], + "description": "OpenVoice VC model from https://huggingface.co/myshell-ai/OpenVoiceV2", + "author": "MyShell.ai", + "license": "MIT", + "commit": null + } } } } diff --git a/TTS/VERSION b/TTS/VERSION deleted file mode 100644 index 2157409059..0000000000 --- a/TTS/VERSION +++ /dev/null @@ -1 +0,0 @@ -0.22.0 diff --git a/TTS/__init__.py b/TTS/__init__.py index eaf05db1b9..8e93c9b5db 100644 --- a/TTS/__init__.py +++ b/TTS/__init__.py @@ -1,6 +1,33 @@ -import os +import importlib.metadata -with open(os.path.join(os.path.dirname(__file__), "VERSION"), "r", encoding="utf-8") as f: - version = f.read().strip() +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 -__version__ = version +__version__ = importlib.metadata.version("coqui-tts") + + +if is_pytorch_at_least_2_4(): + import _codecs + from collections import defaultdict + + import numpy as np + import torch + + from TTS.config.shared_configs import BaseDatasetConfig + from TTS.tts.configs.xtts_config import XttsConfig + from TTS.tts.models.xtts import XttsArgs, XttsAudioConfig + from TTS.utils.radam import RAdam + + torch.serialization.add_safe_globals([dict, defaultdict, RAdam]) + + # Bark + torch.serialization.add_safe_globals( + [ + np.core.multiarray.scalar, + np.dtype, + np.dtypes.Float64DType, + _codecs.encode, # TODO: safe by default from Pytorch 2.5 + ] + ) + + # XTTS + torch.serialization.add_safe_globals([BaseDatasetConfig, XttsConfig, XttsAudioConfig, XttsArgs]) diff --git a/TTS/api.py b/TTS/api.py index 2277ff270c..86a311112e 100644 --- a/TTS/api.py +++ b/TTS/api.py @@ -1,15 +1,18 @@ +"""Coqui TTS Python API.""" + +import logging import tempfile import warnings from pathlib import Path -from typing import Union +from typing import Optional -import numpy as np from torch import nn -from TTS.utils.audio.numpy_transforms import save_wav +from TTS.config import load_config from TTS.utils.manage import ModelManager from TTS.utils.synthesizer import Synthesizer -from TTS.config import load_config + +logger = logging.getLogger(__name__) class TTS(nn.Module): @@ -18,13 +21,19 @@ class TTS(nn.Module): def __init__( self, model_name: str = "", - model_path: str = None, - config_path: str = None, - vocoder_path: str = None, - vocoder_config_path: str = None, + *, + model_path: Optional[str] = None, + config_path: Optional[str] = None, + vocoder_name: Optional[str] = None, + vocoder_path: Optional[str] = None, + vocoder_config_path: Optional[str] = None, + encoder_path: Optional[str] = None, + encoder_config_path: Optional[str] = None, + speakers_file_path: Optional[str] = None, + language_ids_file_path: Optional[str] = None, progress_bar: bool = True, - gpu=False, - ): + gpu: bool = False, + ) -> None: """🐸TTS python interface that allows to load and use the released models. Example with a multi-speaker model: @@ -34,118 +43,147 @@ def __init__( >>> tts.tts_to_file(text="Hello world!", speaker=tts.speakers[0], language=tts.languages[0], file_path="output.wav") Example with a single-speaker model: - >>> tts = TTS(model_name="tts_models/de/thorsten/tacotron2-DDC", progress_bar=False, gpu=False) + >>> tts = TTS(model_name="tts_models/de/thorsten/tacotron2-DDC", progress_bar=False) >>> tts.tts_to_file(text="Ich bin eine Testnachricht.", file_path="output.wav") Example loading a model from a path: - >>> tts = TTS(model_path="/path/to/checkpoint_100000.pth", config_path="/path/to/config.json", progress_bar=False, gpu=False) + >>> tts = TTS(model_path="/path/to/checkpoint_100000.pth", config_path="/path/to/config.json", progress_bar=False) >>> tts.tts_to_file(text="Ich bin eine Testnachricht.", file_path="output.wav") Example voice cloning with YourTTS in English, French and Portuguese: - >>> tts = TTS(model_name="tts_models/multilingual/multi-dataset/your_tts", progress_bar=False, gpu=True) + >>> tts = TTS(model_name="tts_models/multilingual/multi-dataset/your_tts", progress_bar=False).to("cuda") >>> tts.tts_to_file("This is voice cloning.", speaker_wav="my/cloning/audio.wav", language="en", file_path="thisisit.wav") >>> tts.tts_to_file("C'est le clonage de la voix.", speaker_wav="my/cloning/audio.wav", language="fr", file_path="thisisit.wav") >>> tts.tts_to_file("Isso Ê clonagem de voz.", speaker_wav="my/cloning/audio.wav", language="pt", file_path="thisisit.wav") Example Fairseq TTS models (uses ISO language codes in https://dl.fbaipublicfiles.com/mms/tts/all-tts-languages.html): - >>> tts = TTS(model_name="tts_models/eng/fairseq/vits", progress_bar=False, gpu=True) + >>> tts = TTS(model_name="tts_models/eng/fairseq/vits", progress_bar=False).to("cuda") >>> tts.tts_to_file("This is a test.", file_path="output.wav") Args: model_name (str, optional): Model name to load. You can list models by ```tts.models```. Defaults to None. model_path (str, optional): Path to the model checkpoint. Defaults to None. config_path (str, optional): Path to the model config. Defaults to None. + vocoder_name (str, optional): Pre-trained vocoder to use. Defaults to None, i.e. using the default vocoder. vocoder_path (str, optional): Path to the vocoder checkpoint. Defaults to None. vocoder_config_path (str, optional): Path to the vocoder config. Defaults to None. - progress_bar (bool, optional): Whether to pring a progress bar while downloading a model. Defaults to True. - gpu (bool, optional): Enable/disable GPU. Some models might be too slow on CPU. Defaults to False. + encoder_path: Path to speaker encoder checkpoint. Default to None. + encoder_config_path: Path to speaker encoder config file. Defaults to None. + speakers_file_path: JSON file for multi-speaker model. Defaults to None. + language_ids_file_path: JSON file for multilingual model. Defaults to None + progress_bar (bool, optional): Whether to print a progress bar while downloading a model. Defaults to True. + gpu (bool, optional): Enable/disable GPU. Defaults to False. DEPRECATED, use TTS(...).to("cuda") """ super().__init__() - self.manager = ModelManager(models_file=self.get_models_file_path(), progress_bar=progress_bar, verbose=False) + self.manager = ModelManager(models_file=self.get_models_file_path(), progress_bar=progress_bar) self.config = load_config(config_path) if config_path else None self.synthesizer = None self.voice_converter = None self.model_name = "" + + self.vocoder_path = vocoder_path + self.vocoder_config_path = vocoder_config_path + self.encoder_path = encoder_path + self.encoder_config_path = encoder_config_path + self.speakers_file_path = speakers_file_path + self.language_ids_file_path = language_ids_file_path + if gpu: warnings.warn("`gpu` will be deprecated. Please use `tts.to(device)` instead.") if model_name is not None and len(model_name) > 0: if "tts_models" in model_name: - self.load_tts_model_by_name(model_name, gpu) + self.load_tts_model_by_name(model_name, vocoder_name, gpu=gpu) elif "voice_conversion_models" in model_name: - self.load_vc_model_by_name(model_name, gpu) + self.load_vc_model_by_name(model_name, gpu=gpu) + # To allow just TTS("xtts") else: - self.load_model_by_name(model_name, gpu) + self.load_model_by_name(model_name, vocoder_name, gpu=gpu) if model_path: - self.load_tts_model_by_path( - model_path, config_path, vocoder_path=vocoder_path, vocoder_config=vocoder_config_path, gpu=gpu - ) + self.load_tts_model_by_path(model_path, config_path, gpu=gpu) @property - def models(self): + def models(self) -> list[str]: return self.manager.list_tts_models() @property - def is_multi_speaker(self): - if hasattr(self.synthesizer.tts_model, "speaker_manager") and self.synthesizer.tts_model.speaker_manager: + def is_multi_speaker(self) -> bool: + if ( + self.synthesizer is not None + and hasattr(self.synthesizer.tts_model, "speaker_manager") + and self.synthesizer.tts_model.speaker_manager + ): return self.synthesizer.tts_model.speaker_manager.num_speakers > 1 return False @property - def is_multi_lingual(self): + def is_multi_lingual(self) -> bool: # Not sure what sets this to None, but applied a fix to prevent crashing. if ( isinstance(self.model_name, str) and "xtts" in self.model_name or self.config - and ("xtts" in self.config.model or len(self.config.languages) > 1) + and ("xtts" in self.config.model or "languages" in self.config and len(self.config.languages) > 1) ): return True - if hasattr(self.synthesizer.tts_model, "language_manager") and self.synthesizer.tts_model.language_manager: + if ( + self.synthesizer is not None + and hasattr(self.synthesizer.tts_model, "language_manager") + and self.synthesizer.tts_model.language_manager + ): return self.synthesizer.tts_model.language_manager.num_languages > 1 return False @property - def speakers(self): + def speakers(self) -> list[str]: if not self.is_multi_speaker: return None return self.synthesizer.tts_model.speaker_manager.speaker_names @property - def languages(self): + def languages(self) -> list[str]: if not self.is_multi_lingual: return None return self.synthesizer.tts_model.language_manager.language_names @staticmethod - def get_models_file_path(): + def get_models_file_path() -> Path: return Path(__file__).parent / ".models.json" - def list_models(self): - return ModelManager(models_file=TTS.get_models_file_path(), progress_bar=False, verbose=False) + @staticmethod + def list_models() -> list[str]: + return ModelManager(models_file=TTS.get_models_file_path(), progress_bar=False).list_models() - def download_model_by_name(self, model_name: str): + def download_model_by_name( + self, model_name: str, vocoder_name: Optional[str] = None + ) -> tuple[Optional[Path], Optional[Path], Optional[Path]]: model_path, config_path, model_item = self.manager.download_model(model_name) if "fairseq" in model_name or (model_item is not None and isinstance(model_item["model_url"], list)): # return model directory if there are multiple files # we assume that the model knows how to load itself - return None, None, None, None, model_path + return None, None, model_path if model_item.get("default_vocoder") is None: - return model_path, config_path, None, None, None - vocoder_path, vocoder_config_path, _ = self.manager.download_model(model_item["default_vocoder"]) - return model_path, config_path, vocoder_path, vocoder_config_path, None - - def load_model_by_name(self, model_name: str, gpu: bool = False): + return model_path, config_path, None + if vocoder_name is None: + vocoder_name = model_item["default_vocoder"] + vocoder_path, vocoder_config_path, _ = self.manager.download_model(vocoder_name) + # A local vocoder model will take precedence if specified via vocoder_path + if self.vocoder_path is None or self.vocoder_config_path is None: + self.vocoder_path = vocoder_path + self.vocoder_config_path = vocoder_config_path + return model_path, config_path, None + + def load_model_by_name(self, model_name: str, vocoder_name: Optional[str] = None, *, gpu: bool = False) -> None: """Load one of the 🐸TTS models by name. Args: model_name (str): Model name to load. You can list models by ```tts.models```. gpu (bool, optional): Enable/disable GPU. Some models might be too slow on CPU. Defaults to False. """ - self.load_tts_model_by_name(model_name, gpu) + self.load_tts_model_by_name(model_name, vocoder_name, gpu=gpu) - def load_vc_model_by_name(self, model_name: str, gpu: bool = False): + def load_vc_model_by_name(self, model_name: str, *, gpu: bool = False) -> None: """Load one of the voice conversion models by name. Args: @@ -153,10 +191,12 @@ def load_vc_model_by_name(self, model_name: str, gpu: bool = False): gpu (bool, optional): Enable/disable GPU. Some models might be too slow on CPU. Defaults to False. """ self.model_name = model_name - model_path, config_path, _, _, _ = self.download_model_by_name(model_name) - self.voice_converter = Synthesizer(vc_checkpoint=model_path, vc_config=config_path, use_cuda=gpu) + model_path, config_path, model_dir = self.download_model_by_name(model_name) + self.voice_converter = Synthesizer( + vc_checkpoint=model_path, vc_config=config_path, model_dir=model_dir, use_cuda=gpu + ) - def load_tts_model_by_name(self, model_name: str, gpu: bool = False): + def load_tts_model_by_name(self, model_name: str, vocoder_name: Optional[str] = None, *, gpu: bool = False) -> None: """Load one of 🐸TTS models by name. Args: @@ -168,9 +208,7 @@ def load_tts_model_by_name(self, model_name: str, gpu: bool = False): self.synthesizer = None self.model_name = model_name - model_path, config_path, vocoder_path, vocoder_config_path, model_dir = self.download_model_by_name( - model_name - ) + model_path, config_path, model_dir = self.download_model_by_name(model_name, vocoder_name) # init synthesizer # None values are fetch from the model @@ -179,17 +217,15 @@ def load_tts_model_by_name(self, model_name: str, gpu: bool = False): tts_config_path=config_path, tts_speakers_file=None, tts_languages_file=None, - vocoder_checkpoint=vocoder_path, - vocoder_config=vocoder_config_path, - encoder_checkpoint=None, - encoder_config=None, + vocoder_checkpoint=self.vocoder_path, + vocoder_config=self.vocoder_config_path, + encoder_checkpoint=self.encoder_path, + encoder_config=self.encoder_config_path, model_dir=model_dir, use_cuda=gpu, ) - def load_tts_model_by_path( - self, model_path: str, config_path: str, vocoder_path: str = None, vocoder_config: str = None, gpu: bool = False - ): + def load_tts_model_by_path(self, model_path: str, config_path: str, *, gpu: bool = False) -> None: """Load a model from a path. Args: @@ -203,22 +239,22 @@ def load_tts_model_by_path( self.synthesizer = Synthesizer( tts_checkpoint=model_path, tts_config_path=config_path, - tts_speakers_file=None, - tts_languages_file=None, - vocoder_checkpoint=vocoder_path, - vocoder_config=vocoder_config, - encoder_checkpoint=None, - encoder_config=None, + tts_speakers_file=self.speakers_file_path, + tts_languages_file=self.language_ids_file_path, + vocoder_checkpoint=self.vocoder_path, + vocoder_config=self.vocoder_config_path, + encoder_checkpoint=self.encoder_path, + encoder_config=self.encoder_config_path, use_cuda=gpu, ) def _check_arguments( self, - speaker: str = None, - language: str = None, - speaker_wav: str = None, - emotion: str = None, - speed: float = None, + speaker: Optional[str] = None, + language: Optional[str] = None, + speaker_wav: Optional[str] = None, + emotion: Optional[str] = None, + speed: Optional[float] = None, **kwargs, ) -> None: """Check if the arguments are valid for the model.""" @@ -231,7 +267,7 @@ def _check_arguments( raise ValueError("Model is not multi-speaker but `speaker` is provided.") if not self.is_multi_lingual and language is not None: raise ValueError("Model is not multi-lingual but `language` is provided.") - if not emotion is None and not speed is None: + if emotion is not None and speed is not None: raise ValueError("Emotion and speed can only be used with Coqui Studio models. Which is discontinued.") def tts( @@ -278,10 +314,6 @@ def tts( speaker_name=speaker, language_name=language, speaker_wav=speaker_wav, - reference_wav=None, - style_wav=None, - style_text=None, - reference_speaker_name=None, split_sentences=split_sentences, speed=speed, **kwargs, @@ -300,7 +332,7 @@ def tts_to_file( file_path: str = "output.wav", split_sentences: bool = True, **kwargs, - ): + ) -> str: """Convert text to speech. Args: @@ -356,15 +388,18 @@ def voice_conversion( target_wav (str):` Path to the target wav file. """ - wav = self.voice_converter.voice_conversion(source_wav=source_wav, target_wav=target_wav) - return wav + if self.voice_converter is None: + msg = "The selected model does not support voice conversion." + raise RuntimeError(msg) + return self.voice_converter.voice_conversion(source_wav=source_wav, target_wav=target_wav) def voice_conversion_to_file( self, source_wav: str, target_wav: str, file_path: str = "output.wav", - ): + pipe_out=None, + ) -> str: """Voice conversion with FreeVC. Convert source wav to target speaker. Args: @@ -374,9 +409,11 @@ def voice_conversion_to_file( Path to the target wav file. file_path (str, optional): Output file path. Defaults to "output.wav". + pipe_out (BytesIO, optional): + Flag to stdout the generated TTS wav file for shell pipe. """ wav = self.voice_conversion(source_wav=source_wav, target_wav=target_wav) - save_wav(wav=wav, path=file_path, sample_rate=self.voice_converter.vc_config.audio.output_sample_rate) + self.voice_converter.save_wav(wav=wav, path=file_path, pipe_out=pipe_out) return file_path def tts_with_vc( @@ -429,7 +466,8 @@ def tts_with_vc_to_file( file_path: str = "output.wav", speaker: str = None, split_sentences: bool = True, - ): + pipe_out=None, + ) -> str: """Convert text to speech with voice conversion and save to file. Check `tts_with_vc` for more details. @@ -452,8 +490,11 @@ def tts_with_vc_to_file( Split text into sentences, synthesize them separately and concatenate the file audio. Setting it False uses more VRAM and possibly hit model specific text length or VRAM limits. Only applicable to the 🐸TTS models. Defaults to True. + pipe_out (BytesIO, optional): + Flag to stdout the generated TTS wav file for shell pipe. """ wav = self.tts_with_vc( text=text, language=language, speaker_wav=speaker_wav, speaker=speaker, split_sentences=split_sentences ) - save_wav(wav=wav, path=file_path, sample_rate=self.voice_converter.vc_config.audio.output_sample_rate) + self.voice_converter.save_wav(wav=wav, path=file_path, pipe_out=pipe_out) + return file_path diff --git a/TTS/bin/collect_env_info.py b/TTS/bin/collect_env_info.py index 662fcd02ec..32aa303e6e 100644 --- a/TTS/bin/collect_env_info.py +++ b/TTS/bin/collect_env_info.py @@ -1,4 +1,6 @@ """Get detailed info about the working environment.""" + +import json import os import platform import sys @@ -6,11 +8,10 @@ import numpy import torch -sys.path += [os.path.abspath(".."), os.path.abspath(".")] -import json - import TTS +sys.path += [os.path.abspath(".."), os.path.abspath(".")] + def system_info(): return { diff --git a/TTS/bin/compute_attention_masks.py b/TTS/bin/compute_attention_masks.py index 9ab520be7d..b8f69b54e5 100644 --- a/TTS/bin/compute_attention_masks.py +++ b/TTS/bin/compute_attention_masks.py @@ -1,21 +1,26 @@ import argparse import importlib +import logging import os +import sys from argparse import RawTextHelpFormatter import numpy as np import torch from torch.utils.data import DataLoader from tqdm import tqdm +from trainer.io import load_checkpoint from TTS.config import load_config from TTS.tts.datasets.TTSDataset import TTSDataset from TTS.tts.models import setup_model from TTS.tts.utils.text.characters import make_symbols, phonemes, symbols from TTS.utils.audio import AudioProcessor -from TTS.utils.io import load_checkpoint +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger if __name__ == "__main__": + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + # pylint: disable=bad-option-value parser = argparse.ArgumentParser( description="""Extract attention masks from trained Tacotron/Tacotron2 models. @@ -31,7 +36,7 @@ --data_path /root/LJSpeech-1.1/ --batch_size 32 --dataset ljspeech - --use_cuda True + --use_cuda """, formatter_class=RawTextHelpFormatter, ) @@ -58,7 +63,7 @@ help="Dataset metafile inclusing file paths with transcripts.", ) parser.add_argument("--data_path", type=str, default="", help="Defines the data path. It overwrites config.json.") - parser.add_argument("--use_cuda", type=bool, default=False, help="enable/disable cuda.") + parser.add_argument("--use_cuda", action=argparse.BooleanOptionalAction, default=False, help="enable/disable cuda.") parser.add_argument( "--batch_size", default=16, type=int, help="Batch size for the model. Use batch_size=1 if you have no CUDA." @@ -70,13 +75,13 @@ # if the vocabulary was passed, replace the default if "characters" in C.keys(): - symbols, phonemes = make_symbols(**C.characters) + symbols, phonemes = make_symbols(**C.characters) # noqa: F811 # load the model num_chars = len(phonemes) if C.use_phonemes else len(symbols) # TODO: handle multi-speaker model = setup_model(C) - model, _ = load_checkpoint(model, args.model_path, args.use_cuda, True) + model, _ = load_checkpoint(model, args.model_path, use_cuda=args.use_cuda, eval=True) # data loader preprocessor = importlib.import_module("TTS.tts.datasets.formatters") diff --git a/TTS/bin/compute_embeddings.py b/TTS/bin/compute_embeddings.py index 5b5a37df73..dc0ce5b18b 100644 --- a/TTS/bin/compute_embeddings.py +++ b/TTS/bin/compute_embeddings.py @@ -1,5 +1,7 @@ import argparse +import logging import os +import sys from argparse import RawTextHelpFormatter import torch @@ -10,6 +12,7 @@ from TTS.tts.datasets import load_tts_samples from TTS.tts.utils.managers import save_file from TTS.tts.utils.speakers import SpeakerManager +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger def compute_embeddings( @@ -100,6 +103,8 @@ def compute_embeddings( if __name__ == "__main__": + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + parser = argparse.ArgumentParser( description="""Compute embedding vectors for each audio file in a dataset and store them keyed by `{dataset_name}#{file_path}` in a .pth file\n\n""" """ @@ -146,7 +151,7 @@ def compute_embeddings( default=False, action="store_true", ) - parser.add_argument("--disable_cuda", type=bool, help="Flag to disable cuda.", default=False) + parser.add_argument("--disable_cuda", action="store_true", help="Flag to disable cuda.", default=False) parser.add_argument("--no_eval", help="Do not compute eval?. Default False", default=False, action="store_true") parser.add_argument( "--formatter_name", diff --git a/TTS/bin/compute_statistics.py b/TTS/bin/compute_statistics.py index 3ab7ea7a3b..acec91c369 100755 --- a/TTS/bin/compute_statistics.py +++ b/TTS/bin/compute_statistics.py @@ -3,7 +3,9 @@ import argparse import glob +import logging import os +import sys import numpy as np from tqdm import tqdm @@ -12,10 +14,13 @@ from TTS.config import load_config from TTS.tts.datasets import load_tts_samples from TTS.utils.audio import AudioProcessor +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger def main(): """Run preprocessing process.""" + setup_logger("TTS", level=logging.INFO, stream=sys.stderr, formatter=ConsoleFormatter()) + parser = argparse.ArgumentParser(description="Compute mean and variance of spectrogtram features.") parser.add_argument("config_path", type=str, help="TTS config file path to define audio processin parameters.") parser.add_argument("out_path", type=str, help="save path (directory and filename).") diff --git a/TTS/bin/eval_encoder.py b/TTS/bin/eval_encoder.py index 60fed13932..701c7d8e82 100644 --- a/TTS/bin/eval_encoder.py +++ b/TTS/bin/eval_encoder.py @@ -1,4 +1,6 @@ import argparse +import logging +import sys from argparse import RawTextHelpFormatter import torch @@ -7,6 +9,7 @@ from TTS.config import load_config from TTS.tts.datasets import load_tts_samples from TTS.tts.utils.speakers import SpeakerManager +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger def compute_encoder_accuracy(dataset_items, encoder_manager): @@ -51,6 +54,8 @@ def compute_encoder_accuracy(dataset_items, encoder_manager): if __name__ == "__main__": + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + parser = argparse.ArgumentParser( description="""Compute the accuracy of the encoder.\n\n""" """ @@ -71,8 +76,8 @@ def compute_encoder_accuracy(dataset_items, encoder_manager): type=str, help="Path to dataset config file.", ) - parser.add_argument("--use_cuda", type=bool, help="flag to set cuda.", default=True) - parser.add_argument("--eval", type=bool, help="compute eval.", default=True) + parser.add_argument("--use_cuda", action=argparse.BooleanOptionalAction, help="flag to set cuda.", default=True) + parser.add_argument("--eval", action=argparse.BooleanOptionalAction, help="compute eval.", default=True) args = parser.parse_args() diff --git a/TTS/bin/extract_tts_spectrograms.py b/TTS/bin/extract_tts_spectrograms.py index c6048626b3..a04005ce39 100755 --- a/TTS/bin/extract_tts_spectrograms.py +++ b/TTS/bin/extract_tts_spectrograms.py @@ -2,12 +2,15 @@ """Extract Mel spectrograms with teacher forcing.""" import argparse +import logging import os +import sys import numpy as np import torch from torch.utils.data import DataLoader from tqdm import tqdm +from trainer.generic_utils import count_parameters from TTS.config import load_config from TTS.tts.datasets import TTSDataset, load_tts_samples @@ -16,12 +19,12 @@ from TTS.tts.utils.text.tokenizer import TTSTokenizer from TTS.utils.audio import AudioProcessor from TTS.utils.audio.numpy_transforms import quantize -from TTS.utils.generic_utils import count_parameters +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger use_cuda = torch.cuda.is_available() -def setup_loader(ap, r, verbose=False): +def setup_loader(ap, r): tokenizer, _ = TTSTokenizer.init_from_config(c) dataset = TTSDataset( outputs_per_step=r, @@ -37,7 +40,6 @@ def setup_loader(ap, r, verbose=False): phoneme_cache_path=c.phoneme_cache_path, precompute_num_workers=0, use_noise_augment=False, - verbose=verbose, speaker_id_mapping=speaker_manager.name_to_id if c.use_speaker_embedding else None, d_vector_mapping=speaker_manager.embeddings if c.use_d_vector_file else None, ) @@ -257,7 +259,7 @@ def main(args): # pylint: disable=redefined-outer-name print("\n > Model has {} parameters".format(num_params), flush=True) # set r r = 1 if c.model.lower() == "glow_tts" else model.decoder.r - own_loader = setup_loader(ap, r, verbose=True) + own_loader = setup_loader(ap, r) extract_spectrograms( own_loader, @@ -272,6 +274,8 @@ def main(args): # pylint: disable=redefined-outer-name if __name__ == "__main__": + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + parser = argparse.ArgumentParser() parser.add_argument("--config_path", type=str, help="Path to config file for training.", required=True) parser.add_argument("--checkpoint_path", type=str, help="Model file to be restored.", required=True) @@ -279,7 +283,7 @@ def main(args): # pylint: disable=redefined-outer-name parser.add_argument("--debug", default=False, action="store_true", help="Save audio files for debug") parser.add_argument("--save_audio", default=False, action="store_true", help="Save audio files") parser.add_argument("--quantize_bits", type=int, default=0, help="Save quantized audio files if non-zero") - parser.add_argument("--eval", type=bool, help="compute eval.", default=True) + parser.add_argument("--eval", action=argparse.BooleanOptionalAction, help="compute eval.", default=True) args = parser.parse_args() c = load_config(args.config_path) diff --git a/TTS/bin/find_unique_chars.py b/TTS/bin/find_unique_chars.py index ea16974839..7a7fdf5dd4 100644 --- a/TTS/bin/find_unique_chars.py +++ b/TTS/bin/find_unique_chars.py @@ -1,12 +1,18 @@ """Find all the unique characters in a dataset""" + import argparse +import logging +import sys from argparse import RawTextHelpFormatter from TTS.config import load_config -from TTS.tts.datasets import load_tts_samples +from TTS.tts.datasets import find_unique_chars, load_tts_samples +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger def main(): + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + # pylint: disable=bad-option-value parser = argparse.ArgumentParser( description="""Find all the unique characters or phonemes in a dataset.\n\n""" @@ -28,17 +34,7 @@ def main(): ) items = train_items + eval_items - - texts = "".join(item["text"] for item in items) - chars = set(texts) - lower_chars = filter(lambda c: c.islower(), chars) - chars_force_lower = [c.lower() for c in chars] - chars_force_lower = set(chars_force_lower) - - print(f" > Number of unique characters: {len(chars)}") - print(f" > Unique characters: {''.join(sorted(chars))}") - print(f" > Unique lower characters: {''.join(sorted(lower_chars))}") - print(f" > Unique all forced to lower characters: {''.join(sorted(chars_force_lower))}") + find_unique_chars(items) if __name__ == "__main__": diff --git a/TTS/bin/find_unique_phonemes.py b/TTS/bin/find_unique_phonemes.py index 4bd7a78eef..7c68fdb070 100644 --- a/TTS/bin/find_unique_phonemes.py +++ b/TTS/bin/find_unique_phonemes.py @@ -1,6 +1,9 @@ """Find all the unique characters in a dataset""" + import argparse +import logging import multiprocessing +import sys from argparse import RawTextHelpFormatter from tqdm.contrib.concurrent import process_map @@ -8,15 +11,18 @@ from TTS.config import load_config from TTS.tts.datasets import load_tts_samples from TTS.tts.utils.text.phonemizers import Gruut +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger def compute_phonemes(item): text = item["text"] ph = phonemizer.phonemize(text).replace("|", "") - return set(list(ph)) + return set(ph) def main(): + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + # pylint: disable=W0601 global c, phonemizer # pylint: disable=bad-option-value diff --git a/TTS/bin/remove_silence_using_vad.py b/TTS/bin/remove_silence_using_vad.py index a1eaf4c9a7..f9121d7f77 100755 --- a/TTS/bin/remove_silence_using_vad.py +++ b/TTS/bin/remove_silence_using_vad.py @@ -1,12 +1,15 @@ import argparse import glob +import logging import multiprocessing import os import pathlib +import sys import torch from tqdm import tqdm +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger from TTS.utils.vad import get_vad_model_and_utils, remove_silence torch.set_num_threads(1) @@ -75,8 +78,10 @@ def preprocess_audios(): if __name__ == "__main__": + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + parser = argparse.ArgumentParser( - description="python TTS/bin/remove_silence_using_vad.py -i=VCTK-Corpus/ -o=VCTK-Corpus-removed-silence/ -g=wav48_silence_trimmed/*/*_mic1.flac --trim_just_beginning_and_end True" + description="python TTS/bin/remove_silence_using_vad.py -i=VCTK-Corpus/ -o=VCTK-Corpus-removed-silence/ -g=wav48_silence_trimmed/*/*_mic1.flac --trim_just_beginning_and_end" ) parser.add_argument("-i", "--input_dir", type=str, help="Dataset root dir", required=True) parser.add_argument("-o", "--output_dir", type=str, help="Output Dataset dir", default="") @@ -91,20 +96,20 @@ def preprocess_audios(): parser.add_argument( "-t", "--trim_just_beginning_and_end", - type=bool, + action=argparse.BooleanOptionalAction, default=True, - help="If True this script will trim just the beginning and end nonspeech parts. If False all nonspeech parts will be trim. Default True", + help="If True this script will trim just the beginning and end nonspeech parts. If False all nonspeech parts will be trimmed.", ) parser.add_argument( "-c", "--use_cuda", - type=bool, + action=argparse.BooleanOptionalAction, default=False, help="If True use cuda", ) parser.add_argument( "--use_onnx", - type=bool, + action=argparse.BooleanOptionalAction, default=False, help="If True use onnx", ) diff --git a/TTS/bin/synthesize.py b/TTS/bin/synthesize.py index b86252ab67..5d20db6a59 100755 --- a/TTS/bin/synthesize.py +++ b/TTS/bin/synthesize.py @@ -1,147 +1,141 @@ #!/usr/bin/env python3 -# -*- coding: utf-8 -*- + +"""Command line interface.""" import argparse import contextlib +import logging import sys from argparse import RawTextHelpFormatter # pylint: disable=redefined-outer-name, unused-argument -from pathlib import Path +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger + +logger = logging.getLogger(__name__) description = """ -Synthesize speech on command line. +Synthesize speech on the command line. You can either use your trained model or choose a model from the provided list. -If you don't specify any models, then it uses LJSpeech based English model. - -#### Single Speaker Models - - List provided models: + ```sh + tts --list_models ``` - $ tts --list_models - ``` - -- Get model info (for both tts_models and vocoder_models): - - - Query by type/name: - The model_info_by_name uses the name as it from the --list_models. - ``` - $ tts --model_info_by_name "///" - ``` - For example: - ``` - $ tts --model_info_by_name tts_models/tr/common-voice/glow-tts - $ tts --model_info_by_name vocoder_models/en/ljspeech/hifigan_v2 - ``` - - Query by type/idx: - The model_query_idx uses the corresponding idx from --list_models. - - ``` - $ tts --model_info_by_idx "/" - ``` - - For example: - ``` - $ tts --model_info_by_idx tts_models/3 - ``` +- Get model information. Use the names obtained from `--list_models`. + ```sh + tts --model_info_by_name "///" + ``` + For example: + ```sh + tts --model_info_by_name tts_models/tr/common-voice/glow-tts + tts --model_info_by_name vocoder_models/en/ljspeech/hifigan_v2 + ``` - - Query info for model info by full name: - ``` - $ tts --model_info_by_name "///" - ``` +#### Single speaker models -- Run TTS with default models: +- Run TTS with the default model (`tts_models/en/ljspeech/tacotron2-DDC`): - ``` - $ tts --text "Text for TTS" --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" --out_path output/path/speech.wav ``` - Run TTS and pipe out the generated TTS wav file data: - ``` - $ tts --text "Text for TTS" --pipe_out --out_path output/path/speech.wav | aplay + ```sh + tts --text "Text for TTS" --pipe_out --out_path output/path/speech.wav | aplay ``` - Run a TTS model with its default vocoder model: - ``` - $ tts --text "Text for TTS" --model_name "///" --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" \\ + --model_name "///" \\ + --out_path output/path/speech.wav ``` For example: - ``` - $ tts --text "Text for TTS" --model_name "tts_models/en/ljspeech/glow-tts" --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" \\ + --model_name "tts_models/en/ljspeech/glow-tts" \\ + --out_path output/path/speech.wav ``` -- Run with specific TTS and vocoder models from the list: +- Run with specific TTS and vocoder models from the list. Note that not every vocoder is compatible with every TTS model. - ``` - $ tts --text "Text for TTS" --model_name "///" --vocoder_name "///" --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" \\ + --model_name "///" \\ + --vocoder_name "///" \\ + --out_path output/path/speech.wav ``` For example: - ``` - $ tts --text "Text for TTS" --model_name "tts_models/en/ljspeech/glow-tts" --vocoder_name "vocoder_models/en/ljspeech/univnet" --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" \\ + --model_name "tts_models/en/ljspeech/glow-tts" \\ + --vocoder_name "vocoder_models/en/ljspeech/univnet" \\ + --out_path output/path/speech.wav ``` -- Run your own TTS model (Using Griffin-Lim Vocoder): +- Run your own TTS model (using Griffin-Lim Vocoder): - ``` - $ tts --text "Text for TTS" --model_path path/to/model.pth --config_path path/to/config.json --out_path output/path/speech.wav + ```sh + tts --text "Text for TTS" \\ + --model_path path/to/model.pth \\ + --config_path path/to/config.json \\ + --out_path output/path/speech.wav ``` - Run your own TTS and Vocoder models: - ``` - $ tts --text "Text for TTS" --model_path path/to/model.pth --config_path path/to/config.json --out_path output/path/speech.wav - --vocoder_path path/to/vocoder.pth --vocoder_config_path path/to/vocoder_config.json + ```sh + tts --text "Text for TTS" \\ + --model_path path/to/model.pth \\ + --config_path path/to/config.json \\ + --out_path output/path/speech.wav \\ + --vocoder_path path/to/vocoder.pth \\ + --vocoder_config_path path/to/vocoder_config.json ``` -#### Multi-speaker Models +#### Multi-speaker models -- List the available speakers and choose a among them: +- List the available speakers and choose a `` among them: - ``` - $ tts --model_name "//" --list_speaker_idxs + ```sh + tts --model_name "//" --list_speaker_idxs ``` - Run the multi-speaker TTS model with the target speaker ID: - ``` - $ tts --text "Text for TTS." --out_path output/path/speech.wav --model_name "//" --speaker_idx + ```sh + tts --text "Text for TTS." --out_path output/path/speech.wav \\ + --model_name "//" --speaker_idx ``` - Run your own multi-speaker TTS model: - ``` - $ tts --text "Text for TTS" --out_path output/path/speech.wav --model_path path/to/model.pth --config_path path/to/config.json --speakers_file_path path/to/speaker.json --speaker_idx + ```sh + tts --text "Text for TTS" --out_path output/path/speech.wav \\ + --model_path path/to/model.pth --config_path path/to/config.json \\ + --speakers_file_path path/to/speaker.json --speaker_idx ``` -### Voice Conversion Models +#### Voice conversion models -``` -$ tts --out_path output/path/speech.wav --model_name "//" --source_wav --target_wav +```sh +tts --out_path output/path/speech.wav --model_name "//" \\ + --source_wav --target_wav ``` """ -def str2bool(v): - if isinstance(v, bool): - return v - if v.lower() in ("yes", "true", "t", "y", "1"): - return True - if v.lower() in ("no", "false", "f", "n", "0"): - return False - raise argparse.ArgumentTypeError("Boolean value expected.") - - -def main(): +def parse_args() -> argparse.Namespace: + """Parse arguments.""" parser = argparse.ArgumentParser( description=description.replace(" ```\n", ""), formatter_class=RawTextHelpFormatter, @@ -149,10 +143,7 @@ def main(): parser.add_argument( "--list_models", - type=str2bool, - nargs="?", - const=True, - default=False, + action="store_true", help="list available pre-trained TTS and vocoder models.", ) @@ -200,7 +191,7 @@ def main(): default="tts_output.wav", help="Output wav file path.", ) - parser.add_argument("--use_cuda", type=bool, help="Run model on CUDA.", default=False) + parser.add_argument("--use_cuda", action="store_true", help="Run model on CUDA.") parser.add_argument("--device", type=str, help="Device to run model on.", default="cpu") parser.add_argument( "--vocoder_path", @@ -219,12 +210,9 @@ def main(): parser.add_argument( "--pipe_out", help="stdout the generated TTS wav file for shell pipe.", - type=str2bool, - nargs="?", - const=True, - default=False, + action="store_true", ) - + # args for multi-speaker synthesis parser.add_argument("--speakers_file_path", type=str, help="JSON file for multi-speaker model.", default=None) parser.add_argument("--language_ids_file_path", type=str, help="JSON file for multi-lingual model.", default=None) @@ -254,26 +242,14 @@ def main(): parser.add_argument( "--list_speaker_idxs", help="List available speaker ids for the defined multi-speaker model.", - type=str2bool, - nargs="?", - const=True, - default=False, + action="store_true", ) parser.add_argument( "--list_language_idxs", help="List available language ids for the defined multi-lingual model.", - type=str2bool, - nargs="?", - const=True, - default=False, + action="store_true", ) # aux args - parser.add_argument( - "--save_spectogram", - type=bool, - help="If true save raw spectogram for further (vocoder) processing in out_path.", - default=False, - ) parser.add_argument( "--reference_wav", type=str, @@ -288,8 +264,8 @@ def main(): ) parser.add_argument( "--progress_bar", - type=str2bool, - help="If true shows a progress bar for the model download. Defaults to True", + action=argparse.BooleanOptionalAction, + help="Show a progress bar for the model download.", default=True, ) @@ -330,6 +306,14 @@ def main(): ] if not any(check_args): parser.parse_args(["-h"]) + return args + + +def main() -> None: + """Entry point for `tts` command line interface.""" + args = parse_args() + stream = sys.stderr if args.pipe_out else sys.stdout + setup_logger("TTS", level=logging.INFO, stream=stream, formatter=ConsoleFormatter()) pipe_out = sys.stdout if args.pipe_out else None @@ -337,12 +321,9 @@ def main(): # Late-import to make things load faster from TTS.api import TTS from TTS.utils.manage import ModelManager - from TTS.utils.synthesizer import Synthesizer # load model manager - path = Path(__file__).parent / "../.models.json" - manager = ModelManager(path, progress_bar=args.progress_bar) - api = TTS() + manager = ModelManager(models_file=TTS.get_models_file_path(), progress_bar=args.progress_bar) tts_path = None tts_config_path = None @@ -356,12 +337,12 @@ def main(): vc_config_path = None model_dir = None - # CASE1 #list : list pre-trained TTS models + # 1) List pre-trained TTS models if args.list_models: manager.list_models() sys.exit() - # CASE2 #info : model info for pre-trained TTS models + # 2) Info about pre-trained TTS models (without loading a model) if args.model_info_by_idx: model_query = args.model_info_by_idx manager.model_info_by_idx(model_query) @@ -372,122 +353,83 @@ def main(): manager.model_info_by_full_name(model_query_full_name) sys.exit() - # CASE3: load pre-trained model paths - if args.model_name is not None and not args.model_path: - model_path, config_path, model_item = manager.download_model(args.model_name) - # tts model - if model_item["model_type"] == "tts_models": - tts_path = model_path - tts_config_path = config_path - if "default_vocoder" in model_item: - args.vocoder_name = ( - model_item["default_vocoder"] if args.vocoder_name is None else args.vocoder_name - ) - - # voice conversion model - if model_item["model_type"] == "voice_conversion_models": - vc_path = model_path - vc_config_path = config_path - - # tts model with multiple files to be loaded from the directory path - if model_item.get("author", None) == "fairseq" or isinstance(model_item["model_url"], list): - model_dir = model_path - tts_path = None - tts_config_path = None - args.vocoder_name = None - - # load vocoder - if args.vocoder_name is not None and not args.vocoder_path: - vocoder_path, vocoder_config_path, _ = manager.download_model(args.vocoder_name) - - # CASE4: set custom model paths - if args.model_path is not None: - tts_path = args.model_path - tts_config_path = args.config_path - speakers_file_path = args.speakers_file_path - language_ids_file_path = args.language_ids_file_path - - if args.vocoder_path is not None: - vocoder_path = args.vocoder_path - vocoder_config_path = args.vocoder_config_path - - if args.encoder_path is not None: - encoder_path = args.encoder_path - encoder_config_path = args.encoder_config_path - + # 3) Load a model for further info or TTS/VC device = args.device if args.use_cuda: device = "cuda" - - # load models - synthesizer = Synthesizer( - tts_path, - tts_config_path, - speakers_file_path, - language_ids_file_path, - vocoder_path, - vocoder_config_path, - encoder_path, - encoder_config_path, - vc_path, - vc_config_path, - model_dir, - args.voice_dir, + # A local model will take precedence if specified via modeL_path + model_name = args.model_name if args.model_path is None else None + api = TTS( + model_name=model_name, + model_path=args.model_path, + config_path=args.config_path, + vocoder_name=args.vocoder_name, + vocoder_path=args.vocoder_path, + vocoder_config_path=args.vocoder_config_path, + encoder_path=args.encoder_path, + encoder_config_path=args.encoder_config_path, + speakers_file_path=args.speakers_file_path, + language_ids_file_path=args.language_ids_file_path, + progress_bar=args.progress_bar, ).to(device) # query speaker ids of a multi-speaker model. if args.list_speaker_idxs: - print( - " > Available speaker ids: (Set --speaker_idx flag to one of these values to use the multi-speaker model." + if not api.is_multi_speaker: + logger.info("Model only has a single speaker.") + return + logger.info( + "Available speaker ids: (Set --speaker_idx flag to one of these values to use the multi-speaker model." ) - print(synthesizer.tts_model.speaker_manager.name_to_id) + logger.info(api.speakers) return # query langauge ids of a multi-lingual model. if args.list_language_idxs: - print( - " > Available language ids: (Set --language_idx flag to one of these values to use the multi-lingual model." + if not api.is_multi_lingual: + logger.info("Monolingual model.") + return + logger.info( + "Available language ids: (Set --language_idx flag to one of these values to use the multi-lingual model." ) - print(synthesizer.tts_model.language_manager.name_to_id) + logger.info(api.languages) return # check the arguments against a multi-speaker model. - if synthesizer.tts_speakers_file and (not args.speaker_idx and not args.speaker_wav): - print( - " [!] Looks like you use a multi-speaker model. Define `--speaker_idx` to " + if api.is_multi_speaker and (not args.speaker_idx and not args.speaker_wav): + logger.error( + "Looks like you use a multi-speaker model. Define `--speaker_idx` to " "select the target speaker. You can list the available speakers for this model by `--list_speaker_idxs`." ) return # RUN THE SYNTHESIS if args.text: - print(" > Text: {}".format(args.text)) - - # kick it - if tts_path is not None: - wav = synthesizer.tts( - args.text, - speaker_name=args.speaker_idx, - language_name=args.language_idx, + logger.info("Text: %s", args.text) + + if args.text is not None: + api.tts_to_file( + text=args.text, + speaker=args.speaker_idx, + language=args.language_idx, speaker_wav=args.speaker_wav, + pipe_out=pipe_out, + file_path=args.out_path, reference_wav=args.reference_wav, style_wav=args.capacitron_style_wav, style_text=args.capacitron_style_text, reference_speaker_name=args.reference_speaker_idx, + voice_dir=args.voice_dir, ) - elif vc_path is not None: - wav = synthesizer.voice_conversion( + logger.info("Saved TTS output to %s", args.out_path) + elif args.source_wav is not None and args.target_wav is not None: + api.voice_conversion_to_file( source_wav=args.source_wav, target_wav=args.target_wav, + file_path=args.out_path, + pipe_out=pipe_out, ) - elif model_dir is not None: - wav = synthesizer.tts( - args.text, speaker_name=args.speaker_idx, language_name=args.language_idx, speaker_wav=args.speaker_wav - ) - - # save the results - print(" > Saving output to {}".format(args.out_path)) - synthesizer.save_wav(wav, args.out_path, pipe_out=pipe_out) + logger.info("Saved VC output to %s", args.out_path) if __name__ == "__main__": diff --git a/TTS/bin/train_encoder.py b/TTS/bin/train_encoder.py index a32ad00f56..84123d2db3 100644 --- a/TTS/bin/train_encoder.py +++ b/TTS/bin/train_encoder.py @@ -1,13 +1,16 @@ #!/usr/bin/env python3 # -*- coding: utf-8 -*- +import logging import os import sys import time import traceback +import warnings import torch from torch.utils.data import DataLoader +from trainer.generic_utils import count_parameters, remove_experiment_folder from trainer.io import copy_model_files, save_best_model, save_checkpoint from trainer.torch import NoamLR from trainer.trainer_utils import get_optimizer @@ -18,7 +21,7 @@ from TTS.encoder.utils.visual import plot_embeddings from TTS.tts.datasets import load_tts_samples from TTS.utils.audio import AudioProcessor -from TTS.utils.generic_utils import count_parameters, remove_experiment_folder +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger from TTS.utils.samplers import PerfectBatchSampler from TTS.utils.training import check_update @@ -31,7 +34,7 @@ print(" > Number of GPUs: ", num_gpus) -def setup_loader(ap: AudioProcessor, is_val: bool = False, verbose: bool = False): +def setup_loader(ap: AudioProcessor, is_val: bool = False): num_utter_per_class = c.num_utter_per_class if not is_val else c.eval_num_utter_per_class num_classes_in_batch = c.num_classes_in_batch if not is_val else c.eval_num_classes_in_batch @@ -42,7 +45,6 @@ def setup_loader(ap: AudioProcessor, is_val: bool = False, verbose: bool = False voice_len=c.voice_len, num_utter_per_class=num_utter_per_class, num_classes_in_batch=num_classes_in_batch, - verbose=verbose, augmentation_config=c.audio_augmentation if not is_val else None, use_torch_spec=c.model_params.get("use_torch_spec", False), ) @@ -115,11 +117,14 @@ def evaluation(model, criterion, data_loader, global_step): eval_avg_loss = eval_loss / len(data_loader) # save stats dashboard_logger.eval_stats(global_step, {"loss": eval_avg_loss}) - # plot the last batch in the evaluation - figures = { - "UMAP Plot": plot_embeddings(outputs.detach().cpu().numpy(), c.num_classes_in_batch), - } - dashboard_logger.eval_figures(global_step, figures) + try: + # plot the last batch in the evaluation + figures = { + "UMAP Plot": plot_embeddings(outputs.detach().cpu().numpy(), c.num_classes_in_batch), + } + dashboard_logger.eval_figures(global_step, figures) + except ImportError: + warnings.warn("Install the `umap-learn` package to see embedding plots.") return eval_avg_loss @@ -160,9 +165,6 @@ def train(model, optimizer, scheduler, criterion, data_loader, eval_data_loader, loader_time = time.time() - end_time global_step += 1 - # setup lr - if c.lr_decay: - scheduler.step() optimizer.zero_grad() # dispatch data to GPU @@ -181,6 +183,10 @@ def train(model, optimizer, scheduler, criterion, data_loader, eval_data_loader, grad_norm, _ = check_update(model, c.grad_clip) optimizer.step() + # setup lr + if c.lr_decay: + scheduler.step() + step_time = time.time() - start_time epoch_time += step_time @@ -278,9 +284,9 @@ def main(args): # pylint: disable=redefined-outer-name # pylint: disable=redefined-outer-name meta_data_train, meta_data_eval = load_tts_samples(c.datasets, eval_split=True) - train_data_loader, train_classes, map_classid_to_classname = setup_loader(ap, is_val=False, verbose=True) + train_data_loader, train_classes, map_classid_to_classname = setup_loader(ap, is_val=False) if c.run_eval: - eval_data_loader, _, _ = setup_loader(ap, is_val=True, verbose=True) + eval_data_loader, _, _ = setup_loader(ap, is_val=True) else: eval_data_loader = None @@ -316,6 +322,8 @@ def main(args): # pylint: disable=redefined-outer-name if __name__ == "__main__": + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + args, c, OUT_PATH, AUDIO_PATH, c_logger, dashboard_logger = init_training() try: diff --git a/TTS/bin/train_tts.py b/TTS/bin/train_tts.py index bdb4f6f691..e93b1c9d24 100644 --- a/TTS/bin/train_tts.py +++ b/TTS/bin/train_tts.py @@ -1,4 +1,6 @@ +import logging import os +import sys from dataclasses import dataclass, field from trainer import Trainer, TrainerArgs @@ -6,6 +8,7 @@ from TTS.config import load_config, register_config from TTS.tts.datasets import load_tts_samples from TTS.tts.models import setup_model +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger @dataclass @@ -15,6 +18,8 @@ class TrainTTSArgs(TrainerArgs): def main(): """Run `tts` model training directly by a `config.json` file.""" + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + # init trainer args train_args = TrainTTSArgs() parser = train_args.init_argparse(arg_prefix="") diff --git a/TTS/bin/train_vocoder.py b/TTS/bin/train_vocoder.py index 32ecd7bdc3..aa04177068 100644 --- a/TTS/bin/train_vocoder.py +++ b/TTS/bin/train_vocoder.py @@ -1,10 +1,13 @@ +import logging import os +import sys from dataclasses import dataclass, field from trainer import Trainer, TrainerArgs from TTS.config import load_config, register_config from TTS.utils.audio import AudioProcessor +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger from TTS.vocoder.datasets.preprocess import load_wav_data, load_wav_feat_data from TTS.vocoder.models import setup_model @@ -16,6 +19,8 @@ class TrainVocoderArgs(TrainerArgs): def main(): """Run `tts` model training directly by a `config.json` file.""" + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + # init trainer args train_args = TrainVocoderArgs() parser = train_args.init_argparse(arg_prefix="") diff --git a/TTS/bin/tune_wavegrad.py b/TTS/bin/tune_wavegrad.py index 09582cea7c..d05ae14b7f 100644 --- a/TTS/bin/tune_wavegrad.py +++ b/TTS/bin/tune_wavegrad.py @@ -1,5 +1,8 @@ """Search a good noise schedule for WaveGrad for a given number of inference iterations""" + import argparse +import logging +import sys from itertools import product as cartesian_product import numpy as np @@ -9,11 +12,14 @@ from TTS.config import load_config from TTS.utils.audio import AudioProcessor +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger from TTS.vocoder.datasets.preprocess import load_wav_data from TTS.vocoder.datasets.wavegrad_dataset import WaveGradDataset from TTS.vocoder.models import setup_model if __name__ == "__main__": + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) + parser = argparse.ArgumentParser() parser.add_argument("--model_path", type=str, help="Path to model checkpoint.") parser.add_argument("--config_path", type=str, help="Path to model config file.") @@ -54,7 +60,6 @@ return_segments=False, use_noise_augment=False, use_cache=False, - verbose=True, ) loader = DataLoader( dataset, diff --git a/TTS/config/__init__.py b/TTS/config/__init__.py index c5a6dd68e2..e5f40c0296 100644 --- a/TTS/config/__init__.py +++ b/TTS/config/__init__.py @@ -1,7 +1,7 @@ import json import os import re -from typing import Dict +from typing import Any, Dict, Union import fsspec import yaml @@ -17,9 +17,12 @@ def read_json_with_comments(json_path): with fsspec.open(json_path, "r", encoding="utf-8") as f: input_str = f.read() # handle comments but not urls with // - input_str = re.sub(r"(\"(?:[^\"\\]|\\.)*\")|(/\*(?:.|[\\n\\r])*?\*/)|(//.*)", lambda m: m.group(1) or m.group(2) or "", input_str) + input_str = re.sub( + r"(\"(?:[^\"\\]|\\.)*\")|(/\*(?:.|[\\n\\r])*?\*/)|(//.*)", lambda m: m.group(1) or m.group(2) or "", input_str + ) return json.loads(input_str) + def register_config(model_name: str) -> Coqpit: """Find the right config for the given model name. @@ -65,7 +68,7 @@ def _process_model_name(config_dict: Dict) -> str: return model_name -def load_config(config_path: str) -> Coqpit: +def load_config(config_path: Union[str, os.PathLike[Any]]) -> Coqpit: """Import `json` or `yaml` files as TTS configs. First, load the input file as a `dict` and check the model name to find the corresponding Config class. Then initialize the Config. @@ -78,6 +81,7 @@ def load_config(config_path: str) -> Coqpit: Returns: Coqpit: TTS config object. """ + config_path = str(config_path) config_dict = {} ext = os.path.splitext(config_path)[1] if ext in (".yml", ".yaml"): diff --git a/TTS/demos/xtts_ft_demo/requirements.txt b/TTS/demos/xtts_ft_demo/requirements.txt index cb5b16f66e..b58f41c546 100644 --- a/TTS/demos/xtts_ft_demo/requirements.txt +++ b/TTS/demos/xtts_ft_demo/requirements.txt @@ -1,2 +1,2 @@ faster_whisper==0.9.0 -gradio==4.7.1 \ No newline at end of file +gradio==4.7.1 diff --git a/TTS/demos/xtts_ft_demo/utils/formatter.py b/TTS/demos/xtts_ft_demo/utils/formatter.py index 536faa0108..40e8b8ed32 100644 --- a/TTS/demos/xtts_ft_demo/utils/formatter.py +++ b/TTS/demos/xtts_ft_demo/utils/formatter.py @@ -1,23 +1,17 @@ -import os import gc -import torchaudio +import os + import pandas +import torch +import torchaudio from faster_whisper import WhisperModel -from glob import glob - from tqdm import tqdm -import torch -import torchaudio # torch.set_num_threads(1) - from TTS.tts.layers.xtts.tokenizer import multilingual_cleaners torch.set_num_threads(16) - -import os - audio_types = (".wav", ".mp3", ".flac") @@ -25,9 +19,10 @@ def list_audios(basePath, contains=None): # return the set of files that are valid return list_files(basePath, validExts=audio_types, contains=contains) + def list_files(basePath, validExts=None, contains=None): # loop over the directory structure - for (rootDir, dirNames, filenames) in os.walk(basePath): + for rootDir, dirNames, filenames in os.walk(basePath): # loop over the filenames in the current directory for filename in filenames: # if the contains string is not none and the filename does not contain @@ -36,7 +31,7 @@ def list_files(basePath, validExts=None, contains=None): continue # determine the file extension of the current file - ext = filename[filename.rfind("."):].lower() + ext = filename[filename.rfind(".") :].lower() # check to see if the file is an audio and should be processed if validExts is None or ext.endswith(validExts): @@ -44,13 +39,22 @@ def list_files(basePath, validExts=None, contains=None): audioPath = os.path.join(rootDir, filename) yield audioPath -def format_audio_list(audio_files, target_language="en", out_path=None, buffer=0.2, eval_percentage=0.15, speaker_name="coqui", gradio_progress=None): + +def format_audio_list( + audio_files, + target_language="en", + out_path=None, + buffer=0.2, + eval_percentage=0.15, + speaker_name="coqui", + gradio_progress=None, +): audio_total_size = 0 # make sure that ooutput file exists os.makedirs(out_path, exist_ok=True) # Loading Whisper - device = "cuda" if torch.cuda.is_available() else "cpu" + device = "cuda" if torch.cuda.is_available() else "cpu" print("Loading Whisper Model!") asr_model = WhisperModel("large-v2", device=device, compute_type="float16") @@ -69,7 +73,7 @@ def format_audio_list(audio_files, target_language="en", out_path=None, buffer=0 wav = torch.mean(wav, dim=0, keepdim=True) wav = wav.squeeze() - audio_total_size += (wav.size(-1) / sr) + audio_total_size += wav.size(-1) / sr segments, _ = asr_model.transcribe(audio_path, word_timestamps=True, language=target_language) segments = list(segments) @@ -94,7 +98,7 @@ def format_audio_list(audio_files, target_language="en", out_path=None, buffer=0 # get previous sentence end previous_word_end = words_list[word_idx - 1].end # add buffer or get the silence midle between the previous sentence and the current one - sentence_start = max(sentence_start - buffer, (previous_word_end + sentence_start)/2) + sentence_start = max(sentence_start - buffer, (previous_word_end + sentence_start) / 2) sentence = word.word first_word = False @@ -118,19 +122,16 @@ def format_audio_list(audio_files, target_language="en", out_path=None, buffer=0 # Average the current word end and next word start word_end = min((word.end + next_word_start) / 2, word.end + buffer) - + absoulte_path = os.path.join(out_path, audio_file) os.makedirs(os.path.dirname(absoulte_path), exist_ok=True) i += 1 first_word = True - audio = wav[int(sr*sentence_start):int(sr*word_end)].unsqueeze(0) + audio = wav[int(sr * sentence_start) : int(sr * word_end)].unsqueeze(0) # if the audio is too short ignore it (i.e < 0.33 seconds) - if audio.size(-1) >= sr/3: - torchaudio.save(absoulte_path, - audio, - sr - ) + if audio.size(-1) >= sr / 3: + torchaudio.save(absoulte_path, audio, sr) else: continue @@ -140,21 +141,21 @@ def format_audio_list(audio_files, target_language="en", out_path=None, buffer=0 df = pandas.DataFrame(metadata) df = df.sample(frac=1) - num_val_samples = int(len(df)*eval_percentage) + num_val_samples = int(len(df) * eval_percentage) df_eval = df[:num_val_samples] df_train = df[num_val_samples:] - df_train = df_train.sort_values('audio_file') + df_train = df_train.sort_values("audio_file") train_metadata_path = os.path.join(out_path, "metadata_train.csv") df_train.to_csv(train_metadata_path, sep="|", index=False) eval_metadata_path = os.path.join(out_path, "metadata_eval.csv") - df_eval = df_eval.sort_values('audio_file') + df_eval = df_eval.sort_values("audio_file") df_eval.to_csv(eval_metadata_path, sep="|", index=False) # deallocate VRAM and RAM del asr_model, df_train, df_eval, df, metadata gc.collect() - return train_metadata_path, eval_metadata_path, audio_total_size \ No newline at end of file + return train_metadata_path, eval_metadata_path, audio_total_size diff --git a/TTS/demos/xtts_ft_demo/utils/gpt_train.py b/TTS/demos/xtts_ft_demo/utils/gpt_train.py index a98765c3e7..411a9b0dbe 100644 --- a/TTS/demos/xtts_ft_demo/utils/gpt_train.py +++ b/TTS/demos/xtts_ft_demo/utils/gpt_train.py @@ -1,11 +1,12 @@ -import os import gc +import os from trainer import Trainer, TrainerArgs from TTS.config.shared_configs import BaseDatasetConfig from TTS.tts.datasets import load_tts_samples -from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig, XttsAudioConfig +from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig +from TTS.tts.models.xtts import XttsAudioConfig from TTS.utils.manage import ModelManager @@ -25,7 +26,6 @@ def train_gpt(language, num_epochs, batch_size, grad_acumm, train_csv, eval_csv, BATCH_SIZE = batch_size # set here the batch size GRAD_ACUMM_STEPS = grad_acumm # set here the grad accumulation steps - # Define here the dataset that you want to use for the fine-tuning on. config_dataset = BaseDatasetConfig( formatter="coqui", @@ -43,10 +43,9 @@ def train_gpt(language, num_epochs, batch_size, grad_acumm, train_csv, eval_csv, CHECKPOINTS_OUT_PATH = os.path.join(OUT_PATH, "XTTS_v2.0_original_model_files/") os.makedirs(CHECKPOINTS_OUT_PATH, exist_ok=True) - # DVAE files - DVAE_CHECKPOINT_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/dvae.pth" - MEL_NORM_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/mel_stats.pth" + DVAE_CHECKPOINT_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/dvae.pth" + MEL_NORM_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/mel_stats.pth" # Set the path to the downloaded files DVAE_CHECKPOINT = os.path.join(CHECKPOINTS_OUT_PATH, os.path.basename(DVAE_CHECKPOINT_LINK)) @@ -55,13 +54,14 @@ def train_gpt(language, num_epochs, batch_size, grad_acumm, train_csv, eval_csv, # download DVAE files if needed if not os.path.isfile(DVAE_CHECKPOINT) or not os.path.isfile(MEL_NORM_FILE): print(" > Downloading DVAE files!") - ModelManager._download_model_files([MEL_NORM_LINK, DVAE_CHECKPOINT_LINK], CHECKPOINTS_OUT_PATH, progress_bar=True) - + ModelManager._download_model_files( + [MEL_NORM_LINK, DVAE_CHECKPOINT_LINK], CHECKPOINTS_OUT_PATH, progress_bar=True + ) # Download XTTS v2.0 checkpoint if needed - TOKENIZER_FILE_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/vocab.json" - XTTS_CHECKPOINT_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/model.pth" - XTTS_CONFIG_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/config.json" + TOKENIZER_FILE_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/vocab.json" + XTTS_CHECKPOINT_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/model.pth" + XTTS_CONFIG_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/config.json" # XTTS transfer learning parameters: You we need to provide the paths of XTTS model checkpoint that you want to do the fine tuning. TOKENIZER_FILE = os.path.join(CHECKPOINTS_OUT_PATH, os.path.basename(TOKENIZER_FILE_LINK)) # vocab.json file @@ -160,7 +160,7 @@ def train_gpt(language, num_epochs, batch_size, grad_acumm, train_csv, eval_csv, # get the longest text audio file to use as speaker reference samples_len = [len(item["text"].split(" ")) for item in train_samples] - longest_text_idx = samples_len.index(max(samples_len)) + longest_text_idx = samples_len.index(max(samples_len)) speaker_ref = train_samples[longest_text_idx]["audio_file"] trainer_out_path = trainer.output_path diff --git a/TTS/demos/xtts_ft_demo/xtts_demo.py b/TTS/demos/xtts_ft_demo/xtts_demo.py index ebb11f29d1..7ac38ed6ee 100644 --- a/TTS/demos/xtts_ft_demo/xtts_demo.py +++ b/TTS/demos/xtts_ft_demo/xtts_demo.py @@ -1,19 +1,16 @@ import argparse +import logging import os import sys import tempfile +import traceback import gradio as gr -import librosa.display -import numpy as np - -import os import torch import torchaudio -import traceback + from TTS.demos.xtts_ft_demo.utils.formatter import format_audio_list from TTS.demos.xtts_ft_demo.utils.gpt_train import train_gpt - from TTS.tts.configs.xtts_config import XttsConfig from TTS.tts.models.xtts import Xtts @@ -23,7 +20,10 @@ def clear_gpu_cache(): if torch.cuda.is_available(): torch.cuda.empty_cache() + XTTS_MODEL = None + + def load_model(xtts_checkpoint, xtts_config, xtts_vocab): global XTTS_MODEL clear_gpu_cache() @@ -40,17 +40,23 @@ def load_model(xtts_checkpoint, xtts_config, xtts_vocab): print("Model Loaded!") return "Model Loaded!" + def run_tts(lang, tts_text, speaker_audio_file): if XTTS_MODEL is None or not speaker_audio_file: return "You need to run the previous step to load the model !!", None, None - gpt_cond_latent, speaker_embedding = XTTS_MODEL.get_conditioning_latents(audio_path=speaker_audio_file, gpt_cond_len=XTTS_MODEL.config.gpt_cond_len, max_ref_length=XTTS_MODEL.config.max_ref_len, sound_norm_refs=XTTS_MODEL.config.sound_norm_refs) + gpt_cond_latent, speaker_embedding = XTTS_MODEL.get_conditioning_latents( + audio_path=speaker_audio_file, + gpt_cond_len=XTTS_MODEL.config.gpt_cond_len, + max_ref_length=XTTS_MODEL.config.max_ref_len, + sound_norm_refs=XTTS_MODEL.config.sound_norm_refs, + ) out = XTTS_MODEL.inference( text=tts_text, language=lang, gpt_cond_latent=gpt_cond_latent, speaker_embedding=speaker_embedding, - temperature=XTTS_MODEL.config.temperature, # Add custom parameters here + temperature=XTTS_MODEL.config.temperature, # Add custom parameters here length_penalty=XTTS_MODEL.config.length_penalty, repetition_penalty=XTTS_MODEL.config.repetition_penalty, top_k=XTTS_MODEL.config.top_k, @@ -65,9 +71,7 @@ def run_tts(lang, tts_text, speaker_audio_file): return "Speech generated !", out_path, speaker_audio_file - - -# define a logger to redirect +# define a logger to redirect class Logger: def __init__(self, filename="log.out"): self.log_file = filename @@ -85,21 +89,19 @@ def flush(self): def isatty(self): return False + # redirect stdout and stderr to a file sys.stdout = Logger() sys.stderr = sys.stdout # logging.basicConfig(stream=sys.stdout, level=logging.INFO) -import logging + logging.basicConfig( - level=logging.INFO, - format="%(asctime)s [%(levelname)s] %(message)s", - handlers=[ - logging.StreamHandler(sys.stdout) - ] + level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s", handlers=[logging.StreamHandler(sys.stdout)] ) + def read_logs(): sys.stdout.flush() with open(sys.stdout.log_file, "r") as f: @@ -107,12 +109,11 @@ def read_logs(): if __name__ == "__main__": - parser = argparse.ArgumentParser( description="""XTTS fine-tuning demo\n\n""" """ Example runs: - python3 TTS/demos/xtts_ft_demo/xtts_demo.py --port + python3 TTS/demos/xtts_ft_demo/xtts_demo.py --port """, formatter_class=argparse.RawTextHelpFormatter, ) @@ -190,12 +191,11 @@ def read_logs(): "zh", "hu", "ko", - "ja" + "ja", + "hi", ], ) - progress_data = gr.Label( - label="Progress:" - ) + progress_data = gr.Label(label="Progress:") logs = gr.Textbox( label="Logs:", interactive=False, @@ -203,20 +203,30 @@ def read_logs(): demo.load(read_logs, None, logs, every=1) prompt_compute_btn = gr.Button(value="Step 1 - Create dataset") - + def preprocess_dataset(audio_path, language, out_path, progress=gr.Progress(track_tqdm=True)): clear_gpu_cache() out_path = os.path.join(out_path, "dataset") os.makedirs(out_path, exist_ok=True) if audio_path is None: - return "You should provide one or multiple audio files! If you provided it, probably the upload of the files is not finished yet!", "", "" + return ( + "You should provide one or multiple audio files! If you provided it, probably the upload of the files is not finished yet!", + "", + "", + ) else: try: - train_meta, eval_meta, audio_total_size = format_audio_list(audio_path, target_language=language, out_path=out_path, gradio_progress=progress) + train_meta, eval_meta, audio_total_size = format_audio_list( + audio_path, target_language=language, out_path=out_path, gradio_progress=progress + ) except: traceback.print_exc() error = traceback.format_exc() - return f"The data processing was interrupted due an error !! Please check the console to verify the full error message! \n Error summary: {error}", "", "" + return ( + f"The data processing was interrupted due an error !! Please check the console to verify the full error message! \n Error summary: {error}", + "", + "", + ) clear_gpu_cache() @@ -236,7 +246,7 @@ def preprocess_dataset(audio_path, language, out_path, progress=gr.Progress(trac eval_csv = gr.Textbox( label="Eval CSV:", ) - num_epochs = gr.Slider( + num_epochs = gr.Slider( label="Number of epochs:", minimum=1, maximum=100, @@ -264,9 +274,7 @@ def preprocess_dataset(audio_path, language, out_path, progress=gr.Progress(trac step=1, value=args.max_audio_length, ) - progress_train = gr.Label( - label="Progress:" - ) + progress_train = gr.Label(label="Progress:") logs_tts_train = gr.Textbox( label="Logs:", interactive=False, @@ -274,18 +282,41 @@ def preprocess_dataset(audio_path, language, out_path, progress=gr.Progress(trac demo.load(read_logs, None, logs_tts_train, every=1) train_btn = gr.Button(value="Step 2 - Run the training") - def train_model(language, train_csv, eval_csv, num_epochs, batch_size, grad_acumm, output_path, max_audio_length): + def train_model( + language, train_csv, eval_csv, num_epochs, batch_size, grad_acumm, output_path, max_audio_length + ): clear_gpu_cache() if not train_csv or not eval_csv: - return "You need to run the data processing step or manually set `Train CSV` and `Eval CSV` fields !", "", "", "", "" + return ( + "You need to run the data processing step or manually set `Train CSV` and `Eval CSV` fields !", + "", + "", + "", + "", + ) try: # convert seconds to waveform frames max_audio_length = int(max_audio_length * 22050) - config_path, original_xtts_checkpoint, vocab_file, exp_path, speaker_wav = train_gpt(language, num_epochs, batch_size, grad_acumm, train_csv, eval_csv, output_path=output_path, max_audio_length=max_audio_length) + config_path, original_xtts_checkpoint, vocab_file, exp_path, speaker_wav = train_gpt( + language, + num_epochs, + batch_size, + grad_acumm, + train_csv, + eval_csv, + output_path=output_path, + max_audio_length=max_audio_length, + ) except: traceback.print_exc() error = traceback.format_exc() - return f"The training was interrupted due an error !! Please check the console to check the full error message! \n Error summary: {error}", "", "", "", "" + return ( + f"The training was interrupted due an error !! Please check the console to check the full error message! \n Error summary: {error}", + "", + "", + "", + "", + ) # copy original files to avoid parameters changes issues os.system(f"cp {config_path} {exp_path}") @@ -312,9 +343,7 @@ def train_model(language, train_csv, eval_csv, num_epochs, batch_size, grad_acum label="XTTS vocab path:", value="", ) - progress_load = gr.Label( - label="Progress:" - ) + progress_load = gr.Label(label="Progress:") load_btn = gr.Button(value="Step 3 - Load Fine-tuned XTTS model") with gr.Column() as col2: @@ -342,7 +371,8 @@ def train_model(language, train_csv, eval_csv, num_epochs, batch_size, grad_acum "hu", "ko", "ja", - ] + "hi", + ], ) tts_text = gr.Textbox( label="Input Text.", @@ -351,9 +381,7 @@ def train_model(language, train_csv, eval_csv, num_epochs, batch_size, grad_acum tts_btn = gr.Button(value="Step 4 - Inference") with gr.Column() as col3: - progress_gen = gr.Label( - label="Progress:" - ) + progress_gen = gr.Label(label="Progress:") tts_output_audio = gr.Audio(label="Generated Audio.") reference_audio = gr.Audio(label="Reference audio used.") @@ -371,7 +399,6 @@ def train_model(language, train_csv, eval_csv, num_epochs, batch_size, grad_acum ], ) - train_btn.click( fn=train_model, inputs=[ @@ -386,14 +413,10 @@ def train_model(language, train_csv, eval_csv, num_epochs, batch_size, grad_acum ], outputs=[progress_train, xtts_config, xtts_vocab, xtts_checkpoint, speaker_reference_audio], ) - + load_btn.click( fn=load_model, - inputs=[ - xtts_checkpoint, - xtts_config, - xtts_vocab - ], + inputs=[xtts_checkpoint, xtts_config, xtts_vocab], outputs=[progress_load], ) @@ -407,9 +430,4 @@ def train_model(language, train_csv, eval_csv, num_epochs, batch_size, grad_acum outputs=[progress_gen, tts_output_audio, reference_audio], ) - demo.launch( - share=True, - debug=False, - server_port=args.port, - server_name="0.0.0.0" - ) + demo.launch(share=True, debug=False, server_port=args.port, server_name="0.0.0.0") diff --git a/TTS/encoder/README.md b/TTS/encoder/README.md index b38b20052b..9f829c9e2a 100644 --- a/TTS/encoder/README.md +++ b/TTS/encoder/README.md @@ -14,5 +14,5 @@ To run the code, you need to follow the same flow as in TTS. - Define 'config.json' for your needs. Note that, audio parameters should match your TTS model. - Example training call ```python speaker_encoder/train.py --config_path speaker_encoder/config.json --data_path ~/Data/Libri-TTS/train-clean-360``` -- Generate embedding vectors ```python speaker_encoder/compute_embeddings.py --use_cuda true /model/path/best_model.pth model/config/path/config.json dataset/path/ output_path``` . This code parses all .wav files at the given dataset path and generates the same folder structure under the output path with the generated embedding files. +- Generate embedding vectors ```python speaker_encoder/compute_embeddings.py --use_cuda /model/path/best_model.pth model/config/path/config.json dataset/path/ output_path``` . This code parses all .wav files at the given dataset path and generates the same folder structure under the output path with the generated embedding files. - Watch training on Tensorboard as in TTS diff --git a/TTS/encoder/configs/emotion_encoder_config.py b/TTS/encoder/configs/emotion_encoder_config.py index 5eda2671be..1d12325cf2 100644 --- a/TTS/encoder/configs/emotion_encoder_config.py +++ b/TTS/encoder/configs/emotion_encoder_config.py @@ -1,4 +1,4 @@ -from dataclasses import asdict, dataclass +from dataclasses import dataclass from TTS.encoder.configs.base_encoder_config import BaseEncoderConfig diff --git a/TTS/encoder/configs/speaker_encoder_config.py b/TTS/encoder/configs/speaker_encoder_config.py index 6dceb00277..0588527a68 100644 --- a/TTS/encoder/configs/speaker_encoder_config.py +++ b/TTS/encoder/configs/speaker_encoder_config.py @@ -1,4 +1,4 @@ -from dataclasses import asdict, dataclass +from dataclasses import dataclass from TTS.encoder.configs.base_encoder_config import BaseEncoderConfig diff --git a/TTS/encoder/dataset.py b/TTS/encoder/dataset.py index 582b1fe9ca..bb780e3c1d 100644 --- a/TTS/encoder/dataset.py +++ b/TTS/encoder/dataset.py @@ -1,3 +1,4 @@ +import logging import random import torch @@ -5,6 +6,8 @@ from TTS.encoder.utils.generic_utils import AugmentWAV +logger = logging.getLogger(__name__) + class EncoderDataset(Dataset): def __init__( @@ -15,7 +18,6 @@ def __init__( voice_len=1.6, num_classes_in_batch=64, num_utter_per_class=10, - verbose=False, augmentation_config=None, use_torch_spec=None, ): @@ -24,7 +26,6 @@ def __init__( ap (TTS.tts.utils.AudioProcessor): audio processor object. meta_data (list): list of dataset instances. seq_len (int): voice segment length in seconds. - verbose (bool): print diagnostic information. """ super().__init__() self.config = config @@ -33,7 +34,6 @@ def __init__( self.seq_len = int(voice_len * self.sample_rate) self.num_utter_per_class = num_utter_per_class self.ap = ap - self.verbose = verbose self.use_torch_spec = use_torch_spec self.classes, self.items = self.__parse_items() @@ -50,13 +50,12 @@ def __init__( if "gaussian" in augmentation_config.keys(): self.gaussian_augmentation_config = augmentation_config["gaussian"] - if self.verbose: - print("\n > DataLoader initialization") - print(f" | > Classes per Batch: {num_classes_in_batch}") - print(f" | > Number of instances : {len(self.items)}") - print(f" | > Sequence length: {self.seq_len}") - print(f" | > Num Classes: {len(self.classes)}") - print(f" | > Classes: {self.classes}") + logger.info("DataLoader initialization") + logger.info(" | Classes per batch: %d", num_classes_in_batch) + logger.info(" | Number of instances: %d", len(self.items)) + logger.info(" | Sequence length: %d", self.seq_len) + logger.info(" | Number of classes: %d", len(self.classes)) + logger.info(" | Classes: %s", self.classes) def load_wav(self, filename): audio = self.ap.load_wav(filename, sr=self.ap.sample_rate) diff --git a/TTS/encoder/losses.py b/TTS/encoder/losses.py index 5b5aa0fc48..2e27848c31 100644 --- a/TTS/encoder/losses.py +++ b/TTS/encoder/losses.py @@ -1,7 +1,11 @@ +import logging + import torch import torch.nn.functional as F from torch import nn +logger = logging.getLogger(__name__) + # adapted from https://github.com/cvqluu/GE2E-Loss class GE2ELoss(nn.Module): @@ -23,7 +27,7 @@ def __init__(self, init_w=10.0, init_b=-5.0, loss_method="softmax"): self.b = nn.Parameter(torch.tensor(init_b)) self.loss_method = loss_method - print(" > Initialized Generalized End-to-End loss") + logger.info("Initialized Generalized End-to-End loss") assert self.loss_method in ["softmax", "contrast"] @@ -139,7 +143,7 @@ def __init__(self, init_w=10.0, init_b=-5.0): self.b = nn.Parameter(torch.tensor(init_b)) self.criterion = torch.nn.CrossEntropyLoss() - print(" > Initialized Angular Prototypical loss") + logger.info("Initialized Angular Prototypical loss") def forward(self, x, _label=None): """ @@ -177,7 +181,7 @@ def __init__(self, embedding_dim, n_speakers): self.criterion = torch.nn.CrossEntropyLoss() self.fc = nn.Linear(embedding_dim, n_speakers) - print("Initialised Softmax Loss") + logger.info("Initialised Softmax Loss") def forward(self, x, label=None): # reshape for compatibility @@ -212,7 +216,7 @@ def __init__(self, embedding_dim, n_speakers, init_w=10.0, init_b=-5.0): self.softmax = SoftmaxLoss(embedding_dim, n_speakers) self.angleproto = AngleProtoLoss(init_w, init_b) - print("Initialised SoftmaxAnglePrototypical Loss") + logger.info("Initialised SoftmaxAnglePrototypical Loss") def forward(self, x, label=None): """ diff --git a/TTS/encoder/models/base_encoder.py b/TTS/encoder/models/base_encoder.py index 957ea3c4ca..2082019aad 100644 --- a/TTS/encoder/models/base_encoder.py +++ b/TTS/encoder/models/base_encoder.py @@ -1,12 +1,16 @@ +import logging + import numpy as np import torch import torchaudio from coqpit import Coqpit from torch import nn +from trainer.generic_utils import set_partial_state_dict +from trainer.io import load_fsspec from TTS.encoder.losses import AngleProtoLoss, GE2ELoss, SoftmaxAngleProtoLoss -from TTS.utils.generic_utils import set_init_dict -from TTS.utils.io import load_fsspec + +logger = logging.getLogger(__name__) class PreEmphasis(nn.Module): @@ -118,15 +122,15 @@ def load_checkpoint( state = load_fsspec(checkpoint_path, map_location=torch.device("cpu"), cache=cache) try: self.load_state_dict(state["model"]) - print(" > Model fully restored. ") + logger.info("Model fully restored. ") except (KeyError, RuntimeError) as error: # If eval raise the error if eval: raise error - print(" > Partial model initialization.") + logger.info("Partial model initialization.") model_dict = self.state_dict() - model_dict = set_init_dict(model_dict, state["model"], c) + model_dict = set_partial_state_dict(model_dict, state["model"], config) self.load_state_dict(model_dict) del model_dict @@ -135,7 +139,7 @@ def load_checkpoint( try: criterion.load_state_dict(state["criterion"]) except (KeyError, RuntimeError) as error: - print(" > Criterion load ignored because of:", error) + logger.exception("Criterion load ignored because of: %s", error) # instance and load the criterion for the encoder classifier in inference time if ( diff --git a/TTS/encoder/models/lstm.py b/TTS/encoder/models/lstm.py index 51852b5b82..4e0a7523aa 100644 --- a/TTS/encoder/models/lstm.py +++ b/TTS/encoder/models/lstm.py @@ -86,7 +86,7 @@ def forward(self, x, l2_norm=True): - x: :math:`(N, 1, T_{in})` or :math:`(N, D_{spec}, T_{in})` """ with torch.no_grad(): - with torch.cuda.amp.autocast(enabled=False): + with torch.autocast("cuda", enabled=False): if self.use_torch_spec: x.squeeze_(1) x = self.torch_spec(x) diff --git a/TTS/encoder/utils/generic_utils.py b/TTS/encoder/utils/generic_utils.py index 236d6fe937..495b4def5a 100644 --- a/TTS/encoder/utils/generic_utils.py +++ b/TTS/encoder/utils/generic_utils.py @@ -1,4 +1,5 @@ import glob +import logging import os import random @@ -8,6 +9,8 @@ from TTS.encoder.models.lstm import LSTMSpeakerEncoder from TTS.encoder.models.resnet import ResNetSpeakerEncoder +logger = logging.getLogger(__name__) + class AugmentWAV(object): def __init__(self, ap, augmentation_config): @@ -34,12 +37,14 @@ def __init__(self, ap, augmentation_config): # ignore not listed directories if noise_dir not in self.additive_noise_types: continue - if not noise_dir in self.noise_list: + if noise_dir not in self.noise_list: self.noise_list[noise_dir] = [] self.noise_list[noise_dir].append(wav_file) - print( - f" | > Using Additive Noise Augmentation: with {len(additive_files)} audios instances from {self.additive_noise_types}" + logger.info( + "Using Additive Noise Augmentation: with %d audios instances from %s", + len(additive_files), + self.additive_noise_types, ) self.use_rir = False @@ -50,7 +55,7 @@ def __init__(self, ap, augmentation_config): self.rir_files = glob.glob(os.path.join(self.rir_config["rir_path"], "**/*.wav"), recursive=True) self.use_rir = True - print(f" | > Using RIR Noise Augmentation: with {len(self.rir_files)} audios instances") + logger.info("Using RIR Noise Augmentation: with %d audios instances", len(self.rir_files)) self.create_augmentation_global_list() diff --git a/TTS/encoder/utils/prepare_voxceleb.py b/TTS/encoder/utils/prepare_voxceleb.py index b93baf9e60..37619ed0f8 100644 --- a/TTS/encoder/utils/prepare_voxceleb.py +++ b/TTS/encoder/utils/prepare_voxceleb.py @@ -19,15 +19,19 @@ # pylint: disable=too-many-locals, too-many-statements, too-many-arguments, too-many-instance-attributes """ voxceleb 1 & 2 """ +import csv import hashlib +import logging import os import subprocess import sys import zipfile -import pandas import soundfile as sf -from absl import logging + +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger + +logger = logging.getLogger(__name__) SUBSETS = { "vox1_dev_wav": [ @@ -77,14 +81,14 @@ def download_and_extract(directory, subset, urls): zip_filepath = os.path.join(directory, url.split("/")[-1]) if os.path.exists(zip_filepath): continue - logging.info("Downloading %s to %s" % (url, zip_filepath)) + logger.info("Downloading %s to %s" % (url, zip_filepath)) subprocess.call( "wget %s --user %s --password %s -O %s" % (url, USER["user"], USER["password"], zip_filepath), shell=True, ) statinfo = os.stat(zip_filepath) - logging.info("Successfully downloaded %s, size(bytes): %d" % (url, statinfo.st_size)) + logger.info("Successfully downloaded %s, size(bytes): %d" % (url, statinfo.st_size)) # concatenate all parts into zip files if ".zip" not in zip_filepath: @@ -118,9 +122,9 @@ def exec_cmd(cmd): try: retcode = subprocess.call(cmd, shell=True) if retcode < 0: - logging.info(f"Child was terminated by signal {retcode}") + logger.info(f"Child was terminated by signal {retcode}") except OSError as e: - logging.info(f"Execution failed: {e}") + logger.info(f"Execution failed: {e}") retcode = -999 return retcode @@ -134,11 +138,11 @@ def decode_aac_with_ffmpeg(aac_file, wav_file): bool, True if success. """ cmd = f"ffmpeg -i {aac_file} {wav_file}" - logging.info(f"Decoding aac file using command line: {cmd}") + logger.info(f"Decoding aac file using command line: {cmd}") ret = exec_cmd(cmd) if ret != 0: - logging.error(f"Failed to decode aac file with retcode {ret}") - logging.error("Please check your ffmpeg installation.") + logger.error(f"Failed to decode aac file with retcode {ret}") + logger.error("Please check your ffmpeg installation.") return False return True @@ -152,7 +156,7 @@ def convert_audio_and_make_label(input_dir, subset, output_dir, output_file): output_file: the name of the newly generated csv file. e.g. vox1_dev_wav.csv """ - logging.info("Preprocessing audio and label for subset %s" % subset) + logger.info("Preprocessing audio and label for subset %s" % subset) source_dir = os.path.join(input_dir, subset) files = [] @@ -185,9 +189,12 @@ def convert_audio_and_make_label(input_dir, subset, output_dir, output_file): # Write to CSV file which contains four columns: # "wav_filename", "wav_length_ms", "speaker_id", "speaker_name". csv_file_path = os.path.join(output_dir, output_file) - df = pandas.DataFrame(data=files, columns=["wav_filename", "wav_length_ms", "speaker_id", "speaker_name"]) - df.to_csv(csv_file_path, index=False, sep="\t") - logging.info("Successfully generated csv file {}".format(csv_file_path)) + with open(csv_file_path, "w", newline="", encoding="utf-8") as f: + writer = csv.writer(f, delimiter="\t") + writer.writerow(["wav_filename", "wav_length_ms", "speaker_id", "speaker_name"]) + for wav_file in files: + writer.writerow(wav_file) + logger.info("Successfully generated csv file {}".format(csv_file_path)) def processor(directory, subset, force_process): @@ -200,16 +207,16 @@ def processor(directory, subset, force_process): if not force_process and os.path.exists(subset_csv): return subset_csv - logging.info("Downloading and process the voxceleb in %s", directory) - logging.info("Preparing subset %s", subset) + logger.info("Downloading and process the voxceleb in %s", directory) + logger.info("Preparing subset %s", subset) download_and_extract(directory, subset, urls[subset]) convert_audio_and_make_label(directory, subset, directory, subset + ".csv") - logging.info("Finished downloading and processing") + logger.info("Finished downloading and processing") return subset_csv if __name__ == "__main__": - logging.set_verbosity(logging.INFO) + setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) if len(sys.argv) != 4: print("Usage: python prepare_data.py save_directory user password") sys.exit() diff --git a/TTS/encoder/utils/training.py b/TTS/encoder/utils/training.py index ff8f271d80..48629c7a57 100644 --- a/TTS/encoder/utils/training.py +++ b/TTS/encoder/utils/training.py @@ -2,14 +2,14 @@ from dataclasses import dataclass, field from coqpit import Coqpit -from trainer import TrainerArgs, get_last_checkpoint -from trainer.io import copy_model_files +from trainer import TrainerArgs +from trainer.generic_utils import get_experiment_folder_path, get_git_branch +from trainer.io import copy_model_files, get_last_checkpoint from trainer.logging import logger_factory from trainer.logging.console_logger import ConsoleLogger from TTS.config import load_config, register_config from TTS.tts.utils.text.characters import parse_symbols -from TTS.utils.generic_utils import get_experiment_folder_path, get_git_branch @dataclass @@ -29,7 +29,7 @@ def process_args(args, config=None): args (argparse.Namespace or dict like): Parsed input arguments. config (Coqpit): Model config. If none, it is generated from `args`. Defaults to None. Returns: - c (TTS.utils.io.AttrDict): Config paramaters. + c (Coqpit): Config paramaters. out_path (str): Path to save models and logging. audio_path (str): Path to save generated test audios. c_logger (TTS.utils.console_logger.ConsoleLogger): Class that does diff --git a/TTS/encoder/utils/visual.py b/TTS/encoder/utils/visual.py index 6575b86ec2..bfe40605df 100644 --- a/TTS/encoder/utils/visual.py +++ b/TTS/encoder/utils/visual.py @@ -1,7 +1,6 @@ import matplotlib import matplotlib.pyplot as plt import numpy as np -import umap matplotlib.use("Agg") @@ -30,6 +29,10 @@ def plot_embeddings(embeddings, num_classes_in_batch): + try: + import umap + except ImportError as e: + raise ImportError("Package not installed: umap-learn") from e num_utter_per_class = embeddings.shape[0] // num_classes_in_batch # if necessary get just the first 10 classes diff --git a/TTS/model.py b/TTS/model.py index ae6be7b444..779b1775a3 100644 --- a/TTS/model.py +++ b/TTS/model.py @@ -1,5 +1,6 @@ +import os from abc import abstractmethod -from typing import Dict +from typing import Any, Union import torch from coqpit import Coqpit @@ -11,12 +12,12 @@ class BaseTrainerModel(TrainerModel): """BaseTrainerModel model expanding TrainerModel with required functions by 🐸TTS. - Every new 🐸TTS model must inherit it. + Every new Coqui model must inherit it. """ @staticmethod @abstractmethod - def init_from_config(config: Coqpit): + def init_from_config(config: Coqpit) -> "BaseTrainerModel": """Init the model and all its attributes from the given config. Override this depending on your model. @@ -24,7 +25,7 @@ def init_from_config(config: Coqpit): ... @abstractmethod - def inference(self, input: torch.Tensor, aux_input={}) -> Dict: + def inference(self, input: torch.Tensor, aux_input: dict[str, Any] = {}) -> dict[str, Any]: """Forward pass for inference. It must return a dictionary with the main model output and all the auxiliary outputs. The key ```model_outputs``` @@ -45,15 +46,21 @@ def inference(self, input: torch.Tensor, aux_input={}) -> Dict: @abstractmethod def load_checkpoint( - self, config: Coqpit, checkpoint_path: str, eval: bool = False, strict: bool = True, cache=False + self, + config: Coqpit, + checkpoint_path: Union[str, os.PathLike[Any]], + eval: bool = False, + strict: bool = True, + cache: bool = False, ) -> None: - """Load a model checkpoint gile and get ready for training or inference. + """Load a model checkpoint file and get ready for training or inference. Args: config (Coqpit): Model configuration. - checkpoint_path (str): Path to the model checkpoint file. + checkpoint_path (str | os.PathLike): Path to the model checkpoint file. eval (bool, optional): If true, init model for inference else for training. Defaults to False. strict (bool, optional): Match all checkpoint keys to model's keys. Defaults to True. - cache (bool, optional): If True, cache the file locally for subsequent calls. It is cached under `get_user_data_dir()/tts_cache`. Defaults to False. + cache (bool, optional): If True, cache the file locally for subsequent calls. + It is cached under `trainer.io.get_user_data_dir()/tts_cache`. Defaults to False. """ ... diff --git a/TTS/server/README.md b/TTS/server/README.md index 270656c4e3..ae8e38a4e3 100644 --- a/TTS/server/README.md +++ b/TTS/server/README.md @@ -1,5 +1,8 @@ # :frog: TTS demo server -Before you use the server, make sure you [install](https://github.com/coqui-ai/TTS/tree/dev#install-tts)) :frog: TTS properly. Then, you can follow the steps below. +Before you use the server, make sure you +[install](https://github.com/idiap/coqui-ai-TTS/tree/dev#install-tts)) :frog: TTS +properly and install the additional dependencies with `pip install +coqui-tts[server]`. Then, you can follow the steps below. **Note:** If you install :frog:TTS using ```pip```, you can also use the ```tts-server``` end point on the terminal. @@ -12,7 +15,7 @@ Run the server with the official models. ```python TTS/server/server.py --model_name tts_models/en/ljspeech/tacotron2-DCA --vocoder_name vocoder_models/en/ljspeech/multiband-melgan``` Run the server with the official models on a GPU. -```CUDA_VISIBLE_DEVICES="0" python TTS/server/server.py --model_name tts_models/en/ljspeech/tacotron2-DCA --vocoder_name vocoder_models/en/ljspeech/multiband-melgan --use_cuda True``` +```CUDA_VISIBLE_DEVICES="0" python TTS/server/server.py --model_name tts_models/en/ljspeech/tacotron2-DCA --vocoder_name vocoder_models/en/ljspeech/multiband-melgan --use_cuda``` Run the server with a custom models. ```python TTS/server/server.py --tts_checkpoint /path/to/tts/model.pth --tts_config /path/to/tts/config.json --vocoder_checkpoint /path/to/vocoder/model.pth --vocoder_config /path/to/vocoder/config.json``` diff --git a/TTS/server/server.py b/TTS/server/server.py index 6b2141a9aa..6a4642f9a2 100644 --- a/TTS/server/server.py +++ b/TTS/server/server.py @@ -1,7 +1,11 @@ #!flask/bin/python + +"""TTS demo server.""" + import argparse import io import json +import logging import os import sys from pathlib import Path @@ -9,24 +13,26 @@ from typing import Union from urllib.parse import parse_qs -from flask import Flask, render_template, render_template_string, request, send_file +try: + from flask import Flask, render_template, render_template_string, request, send_file +except ImportError as e: + msg = "Server requires requires flask, use `pip install coqui-tts[server]`" + raise ImportError(msg) from e from TTS.config import load_config +from TTS.utils.generic_utils import ConsoleFormatter, setup_logger from TTS.utils.manage import ModelManager from TTS.utils.synthesizer import Synthesizer +logger = logging.getLogger(__name__) +setup_logger("TTS", level=logging.INFO, stream=sys.stdout, formatter=ConsoleFormatter()) -def create_argparser(): - def convert_boolean(x): - return x.lower() in ["true", "1", "yes"] +def create_argparser() -> argparse.ArgumentParser: parser = argparse.ArgumentParser() parser.add_argument( "--list_models", - type=convert_boolean, - nargs="?", - const=True, - default=False, + action="store_true", help="list available pre-trained tts and vocoder models.", ) parser.add_argument( @@ -54,9 +60,13 @@ def convert_boolean(x): parser.add_argument("--vocoder_config_path", type=str, help="Path to vocoder model config file.", default=None) parser.add_argument("--speakers_file_path", type=str, help="JSON file for multi-speaker model.", default=None) parser.add_argument("--port", type=int, default=5002, help="port to listen on.") - parser.add_argument("--use_cuda", type=convert_boolean, default=False, help="true to use CUDA.") - parser.add_argument("--debug", type=convert_boolean, default=False, help="true to enable Flask debug mode.") - parser.add_argument("--show_details", type=convert_boolean, default=False, help="Generate model detail page.") + parser.add_argument("--use_cuda", action=argparse.BooleanOptionalAction, default=False, help="true to use CUDA.") + parser.add_argument( + "--debug", action=argparse.BooleanOptionalAction, default=False, help="true to enable Flask debug mode." + ) + parser.add_argument( + "--show_details", action=argparse.BooleanOptionalAction, default=False, help="Generate model detail page." + ) return parser @@ -66,10 +76,6 @@ def convert_boolean(x): path = Path(__file__).parent / "../.models.json" manager = ModelManager(path) -if args.list_models: - manager.list_models() - sys.exit() - # update in-use models to the specified released models. model_path = None config_path = None @@ -164,17 +170,15 @@ def index(): def details(): if args.config_path is not None and os.path.isfile(args.config_path): model_config = load_config(args.config_path) - else: - if args.model_name is not None: - model_config = load_config(config_path) + elif args.model_name is not None: + model_config = load_config(config_path) if args.vocoder_config_path is not None and os.path.isfile(args.vocoder_config_path): vocoder_config = load_config(args.vocoder_config_path) + elif args.vocoder_name is not None: + vocoder_config = load_config(vocoder_config_path) else: - if args.vocoder_name is not None: - vocoder_config = load_config(vocoder_config_path) - else: - vocoder_config = None + vocoder_config = None return render_template( "details.html", @@ -197,9 +201,9 @@ def tts(): style_wav = request.headers.get("style-wav") or request.values.get("style_wav", "") style_wav = style_wav_uri_to_dict(style_wav) - print(f" > Model input: {text}") - print(f" > Speaker Idx: {speaker_idx}") - print(f" > Language Idx: {language_idx}") + logger.info("Model input: %s", text) + logger.info("Speaker idx: %s", speaker_idx) + logger.info("Language idx: %s", language_idx) wavs = synthesizer.tts(text, speaker_name=speaker_idx, language_name=language_idx, style_wav=style_wav) out = io.BytesIO() synthesizer.save_wav(wavs, out) @@ -243,7 +247,7 @@ def mary_tts_api_process(): text = data.get("INPUT_TEXT", [""])[0] else: text = request.args.get("INPUT_TEXT", "") - print(f" > Model input: {text}") + logger.info("Model input: %s", text) wavs = synthesizer.tts(text) out = io.BytesIO() synthesizer.save_wav(wavs, out) diff --git a/TTS/server/templates/details.html b/TTS/server/templates/details.html index 51c9ed85a8..85ff959591 100644 --- a/TTS/server/templates/details.html +++ b/TTS/server/templates/details.html @@ -128,4 +128,4 @@ - \ No newline at end of file + diff --git a/TTS/server/templates/index.html b/TTS/server/templates/index.html index 6354d3919d..6bfd5ae2cb 100644 --- a/TTS/server/templates/index.html +++ b/TTS/server/templates/index.html @@ -30,7 +30,7 @@ - Fork me on GitHub @@ -151,4 +151,4 @@ - \ No newline at end of file + diff --git a/TTS/tts/configs/bark_config.py b/TTS/tts/configs/bark_config.py index 4d1cd1374a..b846febe85 100644 --- a/TTS/tts/configs/bark_config.py +++ b/TTS/tts/configs/bark_config.py @@ -2,11 +2,12 @@ from dataclasses import dataclass, field from typing import Dict +from trainer.io import get_user_data_dir + from TTS.tts.configs.shared_configs import BaseTTSConfig from TTS.tts.layers.bark.model import GPTConfig from TTS.tts.layers.bark.model_fine import FineGPTConfig from TTS.tts.models.bark import BarkAudioConfig -from TTS.utils.generic_utils import get_user_data_dir @dataclass @@ -95,7 +96,6 @@ def __post_init__(self): "coarse": os.path.join(self.CACHE_DIR, "coarse_2.pt"), "fine": os.path.join(self.CACHE_DIR, "fine_2.pt"), "hubert_tokenizer": os.path.join(self.CACHE_DIR, "tokenizer.pth"), - "hubert": os.path.join(self.CACHE_DIR, "hubert.pt"), } self.SMALL_REMOTE_MODEL_PATHS = { "text": {"path": os.path.join(self.REMOTE_BASE_URL, "text.pt")}, diff --git a/TTS/tts/datasets/__init__.py b/TTS/tts/datasets/__init__.py index 192138561f..d1a37da4c1 100644 --- a/TTS/tts/datasets/__init__.py +++ b/TTS/tts/datasets/__init__.py @@ -1,3 +1,4 @@ +import logging import os import sys from collections import Counter @@ -9,6 +10,8 @@ from TTS.tts.datasets.dataset import * from TTS.tts.datasets.formatters import * +logger = logging.getLogger(__name__) + def split_dataset(items, eval_split_max_size=None, eval_split_size=0.01): """Split a dataset into train and eval. Consider speaker distribution in multi-speaker training. @@ -122,7 +125,7 @@ def load_tts_samples( meta_data_train = add_extra_keys(meta_data_train, language, dataset_name) - print(f" | > Found {len(meta_data_train)} files in {Path(root_path).resolve()}") + logger.info("Found %d files in %s", len(meta_data_train), Path(root_path).resolve()) # load evaluation split if set if eval_split: if meta_file_val: @@ -163,19 +166,23 @@ def load_attention_mask_meta_data(metafile_path): def _get_formatter_by_name(name): """Returns the respective preprocessing function.""" thismodule = sys.modules[__name__] + if not hasattr(thismodule, name.lower()): + msg = ( + f"{name} formatter not found. If it is a custom formatter, pass the function to load_tts_samples() instead." + ) + raise ValueError(msg) return getattr(thismodule, name.lower()) -def find_unique_chars(data_samples, verbose=True): - texts = "".join(item[0] for item in data_samples) +def find_unique_chars(data_samples): + texts = "".join(item["text"] for item in data_samples) chars = set(texts) lower_chars = filter(lambda c: c.islower(), chars) chars_force_lower = [c.lower() for c in chars] chars_force_lower = set(chars_force_lower) - if verbose: - print(f" > Number of unique characters: {len(chars)}") - print(f" > Unique characters: {''.join(sorted(chars))}") - print(f" > Unique lower characters: {''.join(sorted(lower_chars))}") - print(f" > Unique all forced to lower characters: {''.join(sorted(chars_force_lower))}") + logger.info("Number of unique characters: %d", len(chars)) + logger.info("Unique characters: %s", "".join(sorted(chars))) + logger.info("Unique lower characters: %s", "".join(sorted(lower_chars))) + logger.info("Unique all forced to lower characters: %s", "".join(sorted(chars_force_lower))) return chars_force_lower diff --git a/TTS/tts/datasets/dataset.py b/TTS/tts/datasets/dataset.py index 19fb25bef8..5f629f32a9 100644 --- a/TTS/tts/datasets/dataset.py +++ b/TTS/tts/datasets/dataset.py @@ -1,11 +1,14 @@ import base64 import collections +import logging import os import random -from typing import Dict, List, Union +from typing import Any, Optional, Union import numpy as np +import numpy.typing as npt import torch +import torchaudio import tqdm from torch.utils.data import Dataset @@ -13,7 +16,7 @@ from TTS.utils.audio import AudioProcessor from TTS.utils.audio.numpy_transforms import compute_energy as calculate_energy -import mutagen +logger = logging.getLogger(__name__) # to prevent too many open files error as suggested here # https://github.com/pytorch/pytorch/issues/11201#issuecomment-421146936 @@ -30,27 +33,59 @@ def _parse_sample(item): elif len(item) == 3: text, wav_file, speaker_name = item else: - raise ValueError(" [!] Dataset cannot parse the sample.") + msg = "Dataset cannot parse the sample." + raise ValueError(msg) return text, wav_file, speaker_name, language_name, attn_file -def noise_augment_audio(wav): +def noise_augment_audio(wav: npt.NDArray) -> npt.NDArray: return wav + (1.0 / 32768.0) * np.random.rand(*wav.shape) -def string2filename(string): +def string2filename(string: str) -> str: # generate a safe and reversible filename based on a string - filename = base64.urlsafe_b64encode(string.encode("utf-8")).decode("utf-8", "ignore") - return filename + return base64.urlsafe_b64encode(string.encode("utf-8")).decode("utf-8", "ignore") -def get_audio_size(audiopath): +def get_audio_size(audiopath: Union[str, os.PathLike[Any]]) -> int: + """Return the number of samples in the audio file.""" + if not isinstance(audiopath, str): + audiopath = str(audiopath) extension = audiopath.rpartition(".")[-1].lower() if extension not in {"mp3", "wav", "flac"}: - raise RuntimeError(f"The audio format {extension} is not supported, please convert the audio files to mp3, flac, or wav format!") - - audio_info = mutagen.File(audiopath).info - return int(audio_info.length * audio_info.sample_rate) + msg = f"The audio format {extension} is not supported, please convert the audio files to mp3, flac, or wav format!" + raise RuntimeError(msg) + + try: + return torchaudio.info(audiopath).num_frames + except RuntimeError as e: + msg = f"Failed to decode {audiopath}" + raise RuntimeError(msg) from e + + +def get_attribute_balancer_weights(items: list, attr_name: str, multi_dict: Optional[dict] = None): + """Create inverse frequency weights for balancing the dataset. + + Use `multi_dict` to scale relative weights.""" + attr_names_samples = np.array([item[attr_name] for item in items]) + unique_attr_names = np.unique(attr_names_samples).tolist() + attr_idx = [unique_attr_names.index(l) for l in attr_names_samples] + attr_count = np.array([len(np.where(attr_names_samples == l)[0]) for l in unique_attr_names]) + weight_attr = 1.0 / attr_count + dataset_samples_weight = np.array([weight_attr[l] for l in attr_idx]) + dataset_samples_weight = dataset_samples_weight / np.linalg.norm(dataset_samples_weight) + if multi_dict is not None: + # check if all keys are in the multi_dict + for k in multi_dict: + assert k in unique_attr_names, f"{k} not in {unique_attr_names}" + # scale weights + multiplier_samples = np.array([multi_dict.get(item[attr_name], 1.0) for item in items]) + dataset_samples_weight *= multiplier_samples + return ( + torch.from_numpy(dataset_samples_weight).float(), + unique_attr_names, + np.unique(dataset_samples_weight).tolist(), + ) class TTSDataset(Dataset): @@ -59,32 +94,32 @@ def __init__( outputs_per_step: int = 1, compute_linear_spec: bool = False, ap: AudioProcessor = None, - samples: List[Dict] = None, + samples: Optional[list[dict]] = None, tokenizer: "TTSTokenizer" = None, compute_f0: bool = False, compute_energy: bool = False, - f0_cache_path: str = None, - energy_cache_path: str = None, + f0_cache_path: Optional[str] = None, + energy_cache_path: Optional[str] = None, return_wav: bool = False, batch_group_size: int = 0, min_text_len: int = 0, max_text_len: int = float("inf"), min_audio_len: int = 0, max_audio_len: int = float("inf"), - phoneme_cache_path: str = None, + phoneme_cache_path: Optional[str] = None, precompute_num_workers: int = 0, - speaker_id_mapping: Dict = None, - d_vector_mapping: Dict = None, - language_id_mapping: Dict = None, + speaker_id_mapping: Optional[dict] = None, + d_vector_mapping: Optional[dict] = None, + language_id_mapping: Optional[dict] = None, use_noise_augment: bool = False, start_by_longest: bool = False, - verbose: bool = False, - ): + ) -> None: """Generic 📂 data loader for `tts` models. It is configurable for different outputs and needs. If you need something different, you can subclass and override. Args: + ---- outputs_per_step (int): Number of time frames predicted per step. compute_linear_spec (bool): compute linear spectrogram if True. @@ -137,7 +172,6 @@ def __init__( start_by_longest (bool): Start by longest sequence. It is especially useful to check OOM. Defaults to False. - verbose (bool): Print diagnostic information. Defaults to false. """ super().__init__() self.batch_group_size = batch_group_size @@ -161,33 +195,44 @@ def __init__( self.use_noise_augment = use_noise_augment self.start_by_longest = start_by_longest - self.verbose = verbose self.rescue_item_idx = 1 self.pitch_computed = False self.tokenizer = tokenizer if self.tokenizer.use_phonemes: self.phoneme_dataset = PhonemeDataset( - self.samples, self.tokenizer, phoneme_cache_path, precompute_num_workers=precompute_num_workers + self.samples, + self.tokenizer, + phoneme_cache_path, + precompute_num_workers=precompute_num_workers, ) if compute_f0: self.f0_dataset = F0Dataset( - self.samples, self.ap, cache_path=f0_cache_path, precompute_num_workers=precompute_num_workers + self.samples, + self.ap, + cache_path=f0_cache_path, + precompute_num_workers=precompute_num_workers, ) if compute_energy: self.energy_dataset = EnergyDataset( - self.samples, self.ap, cache_path=energy_cache_path, precompute_num_workers=precompute_num_workers + self.samples, + self.ap, + cache_path=energy_cache_path, + precompute_num_workers=precompute_num_workers, ) - if self.verbose: - self.print_logs() + self.print_logs() @property - def lengths(self): + def lengths(self) -> list[int]: lens = [] for item in self.samples: _, wav_file, *_ = _parse_sample(item) - audio_len = get_audio_size(wav_file) + try: + audio_len = get_audio_size(wav_file) + except RuntimeError: + logger.warning(f"Failed to compute length for {item['audio_file']}") + audio_len = 0 lens.append(audio_len) return lens @@ -196,7 +241,7 @@ def samples(self): return self._samples @samples.setter - def samples(self, new_samples): + def samples(self, new_samples) -> None: self._samples = new_samples if hasattr(self, "f0_dataset"): self.f0_dataset.samples = new_samples @@ -205,7 +250,7 @@ def samples(self, new_samples): if hasattr(self, "phoneme_dataset"): self.phoneme_dataset.samples = new_samples - def __len__(self): + def __len__(self) -> int: return len(self.samples) def __getitem__(self, idx): @@ -213,11 +258,10 @@ def __getitem__(self, idx): def print_logs(self, level: int = 0) -> None: indent = "\t" * level - print("\n") - print(f"{indent}> DataLoader initialization") - print(f"{indent}| > Tokenizer:") + logger.info("%sDataLoader initialization", indent) + logger.info("%s| Tokenizer:", indent) self.tokenizer.print_logs(level + 1) - print(f"{indent}| > Number of instances : {len(self.samples)}") + logger.info("%s| Number of instances : %d", indent, len(self.samples)) def load_wav(self, filename): waveform = self.ap.load_wav(filename) @@ -253,7 +297,7 @@ def get_token_ids(self, idx, text): token_ids = self.tokenizer.text_to_ids(text) return np.array(token_ids, dtype=np.int32) - def load_data(self, idx): + def load_data(self, idx) -> dict[str, Any]: item = self.samples[idx] raw_text = item["text"] @@ -287,7 +331,7 @@ def load_data(self, idx): if self.compute_energy: energy = self.get_energy(idx)["energy"] - sample = { + return { "raw_text": raw_text, "token_ids": token_ids, "wav": wav, @@ -300,13 +344,16 @@ def load_data(self, idx): "wav_file_name": os.path.basename(item["audio_file"]), "audio_unique_name": item["audio_unique_name"], } - return sample @staticmethod def _compute_lengths(samples): new_samples = [] for item in samples: - audio_length = get_audio_size(item["audio_file"]) + try: + audio_length = get_audio_size(item["audio_file"]) + except RuntimeError: + logger.warning(f"Failed to compute length, skipping {item['audio_file']}") + continue text_lenght = len(item["text"]) item["audio_length"] = audio_length item["text_length"] = text_lenght @@ -314,7 +361,7 @@ def _compute_lengths(samples): return new_samples @staticmethod - def filter_by_length(lengths: List[int], min_len: int, max_len: int): + def filter_by_length(lengths: list[int], min_len: int, max_len: int): idxs = np.argsort(lengths) # ascending order ignore_idx = [] keep_idx = [] @@ -327,10 +374,9 @@ def filter_by_length(lengths: List[int], min_len: int, max_len: int): return ignore_idx, keep_idx @staticmethod - def sort_by_length(samples: List[List]): + def sort_by_length(samples: list[list]): audio_lengths = [s["audio_length"] for s in samples] - idxs = np.argsort(audio_lengths) # ascending order - return idxs + return np.argsort(audio_lengths) # ascending order @staticmethod def create_buckets(samples, batch_group_size: int): @@ -350,7 +396,7 @@ def _select_samples_by_idx(idxs, samples): samples_new.append(samples[idx]) return samples_new - def preprocess_samples(self): + def preprocess_samples(self) -> None: r"""Sort `items` based on text length or audio length in ascending order. Filter out samples out or the length range. """ @@ -376,7 +422,8 @@ def preprocess_samples(self): samples = self._select_samples_by_idx(sorted_idxs, samples) if len(samples) == 0: - raise RuntimeError(" [!] No samples left") + msg = "No samples left." + raise RuntimeError(msg) # shuffle batch groups # create batches with similar length items @@ -389,39 +436,38 @@ def preprocess_samples(self): text_lengths = [s["text_length"] for s in samples] self.samples = samples - if self.verbose: - print(" | > Preprocessing samples") - print(" | > Max text length: {}".format(np.max(text_lengths))) - print(" | > Min text length: {}".format(np.min(text_lengths))) - print(" | > Avg text length: {}".format(np.mean(text_lengths))) - print(" | ") - print(" | > Max audio length: {}".format(np.max(audio_lengths))) - print(" | > Min audio length: {}".format(np.min(audio_lengths))) - print(" | > Avg audio length: {}".format(np.mean(audio_lengths))) - print(f" | > Num. instances discarded samples: {len(ignore_idx)}") - print(" | > Batch group size: {}.".format(self.batch_group_size)) + logger.info("Preprocessing samples") + logger.info(f"Max text length: {np.max(text_lengths)}") + logger.info(f"Min text length: {np.min(text_lengths)}") + logger.info(f"Avg text length: {np.mean(text_lengths)}") + logger.info(f"Max audio length: {np.max(audio_lengths)}") + logger.info(f"Min audio length: {np.min(audio_lengths)}") + logger.info(f"Avg audio length: {np.mean(audio_lengths)}") + logger.info("Num. instances discarded samples: %d", len(ignore_idx)) + logger.info(f"Batch group size: {self.batch_group_size}.") @staticmethod def _sort_batch(batch, text_lengths): """Sort the batch by the input text length for RNN efficiency. Args: + ---- batch (Dict): Batch returned by `__getitem__`. text_lengths (List[int]): Lengths of the input character sequences. + """ text_lengths, ids_sorted_decreasing = torch.sort(torch.LongTensor(text_lengths), dim=0, descending=True) batch = [batch[idx] for idx in ids_sorted_decreasing] return batch, text_lengths, ids_sorted_decreasing def collate_fn(self, batch): - r""" - Perform preprocessing and create a final data batch: + """Perform preprocessing and create a final data batch. + 1. Sort batch instances by text-length 2. Convert Audio signal to features. 3. PAD sequences wrt r. 4. Load to Torch. """ - # Puts each data field into a tensor with outer dimension batch size if isinstance(batch[0], collections.abc.Mapping): token_ids_lengths = np.array([len(d["token_ids"]) for d in batch]) @@ -456,9 +502,11 @@ def collate_fn(self, batch): # lengths adjusted by the reduction factor mel_lengths_adjusted = [ - m.shape[1] + (self.outputs_per_step - (m.shape[1] % self.outputs_per_step)) - if m.shape[1] % self.outputs_per_step - else m.shape[1] + ( + m.shape[1] + (self.outputs_per_step - (m.shape[1] % self.outputs_per_step)) + if m.shape[1] % self.outputs_per_step + else m.shape[1] + ) for m in mel ] @@ -564,23 +612,18 @@ def collate_fn(self, batch): "audio_unique_names": batch["audio_unique_name"], } - raise TypeError( - ( - "batch must contain tensors, numbers, dicts or lists;\ - found {}".format( - type(batch[0]) - ) - ) - ) + msg = f"batch must contain tensors, numbers, dicts or lists; found {type(batch[0])}" + raise TypeError(msg) class PhonemeDataset(Dataset): - """Phoneme Dataset for converting input text to phonemes and then token IDs + """Phoneme Dataset for converting input text to phonemes and then token IDs. At initialization, it pre-computes the phonemes under `cache_path` and loads them in training to reduce data loading latency. If `cache_path` is already present, it skips the pre-computation. Args: + ---- samples (Union[List[List], List[Dict]]): List of samples. Each sample is a list or a dict. @@ -592,15 +635,16 @@ class PhonemeDataset(Dataset): precompute_num_workers (int): Number of workers used for pre-computing the phonemes. Defaults to 0. + """ def __init__( self, - samples: Union[List[Dict], List[List]], + samples: Union[list[dict], list[list]], tokenizer: "TTSTokenizer", cache_path: str, - precompute_num_workers=0, - ): + precompute_num_workers: int = 0, + ) -> None: self.samples = samples self.tokenizer = tokenizer self.cache_path = cache_path @@ -608,16 +652,16 @@ def __init__( os.makedirs(cache_path) self.precompute(precompute_num_workers) - def __getitem__(self, index): + def __getitem__(self, index) -> dict[str, Any]: item = self.samples[index] ids = self.compute_or_load(string2filename(item["audio_unique_name"]), item["text"], item["language"]) ph_hat = self.tokenizer.ids_to_text(ids) return {"text": item["text"], "ph_hat": ph_hat, "token_ids": ids, "token_ids_len": len(ids)} - def __len__(self): + def __len__(self) -> int: return len(self.samples) - def compute_or_load(self, file_name, text, language): + def compute_or_load(self, file_name: str, text: str, language: str) -> list[int]: """Compute phonemes for the given text. If the phonemes are already cached, load them from cache. @@ -631,20 +675,24 @@ def compute_or_load(self, file_name, text, language): np.save(cache_path, ids) return ids - def get_pad_id(self): - """Get pad token ID for sequence padding""" + def get_pad_id(self) -> int: + """Get pad token ID for sequence padding.""" return self.tokenizer.pad_id - def precompute(self, num_workers=1): + def precompute(self, num_workers: int = 1) -> None: """Precompute phonemes for all samples. We use pytorch dataloader because we are lazy. """ - print("[*] Pre-computing phonemes...") + logger.info("Pre-computing phonemes...") with tqdm.tqdm(total=len(self)) as pbar: batch_size = num_workers if num_workers > 0 else 1 dataloder = torch.utils.data.DataLoader( - batch_size=batch_size, dataset=self, shuffle=False, num_workers=num_workers, collate_fn=self.collate_fn + batch_size=batch_size, + dataset=self, + shuffle=False, + num_workers=num_workers, + collate_fn=self.collate_fn, ) for _ in dataloder: pbar.update(batch_size) @@ -662,20 +710,20 @@ def collate_fn(self, batch): def print_logs(self, level: int = 0) -> None: indent = "\t" * level - print("\n") - print(f"{indent}> PhonemeDataset ") - print(f"{indent}| > Tokenizer:") + logger.info("%sPhonemeDataset", indent) + logger.info("%s| Tokenizer:", indent) self.tokenizer.print_logs(level + 1) - print(f"{indent}| > Number of instances : {len(self.samples)}") + logger.info("%s| Number of instances : %d", indent, len(self.samples)) class F0Dataset: - """F0 Dataset for computing F0 from wav files in CPU + """F0 Dataset for computing F0 from wav files in CPU. Pre-compute F0 values for all the samples at initialization if `cache_path` is not None or already present. It also computes the mean and std of F0 values if `normalize_f0` is True. Args: + ---- samples (Union[List[List], List[Dict]]): List of samples. Each sample is a list or a dict. @@ -691,21 +739,20 @@ class F0Dataset: normalize_f0 (bool): Whether to normalize F0 values by mean and std. Defaults to True. + """ def __init__( self, - samples: Union[List[List], List[Dict]], + samples: Union[list[list], list[dict]], ap: "AudioProcessor", audio_config=None, # pylint: disable=unused-argument - verbose=False, - cache_path: str = None, - precompute_num_workers=0, - normalize_f0=True, - ): + cache_path: Optional[str] = None, + precompute_num_workers: int = 0, + normalize_f0: bool = True, + ) -> None: self.samples = samples self.ap = ap - self.verbose = verbose self.cache_path = cache_path self.normalize_f0 = normalize_f0 self.pad_id = 0.0 @@ -725,18 +772,22 @@ def __getitem__(self, idx): f0 = self.normalize(f0) return {"audio_unique_name": item["audio_unique_name"], "f0": f0} - def __len__(self): + def __len__(self) -> int: return len(self.samples) - def precompute(self, num_workers=0): - print("[*] Pre-computing F0s...") + def precompute(self, num_workers: int = 0) -> None: + logger.info("Pre-computing F0s...") with tqdm.tqdm(total=len(self)) as pbar: batch_size = num_workers if num_workers > 0 else 1 # we do not normalize at preproessing normalize_f0 = self.normalize_f0 self.normalize_f0 = False dataloder = torch.utils.data.DataLoader( - batch_size=batch_size, dataset=self, shuffle=False, num_workers=num_workers, collate_fn=self.collate_fn + batch_size=batch_size, + dataset=self, + shuffle=False, + num_workers=num_workers, + collate_fn=self.collate_fn, ) computed_data = [] for batch in dataloder: @@ -755,9 +806,8 @@ def get_pad_id(self): return self.pad_id @staticmethod - def create_pitch_file_path(file_name, cache_path): - pitch_file = os.path.join(cache_path, file_name + "_pitch.npy") - return pitch_file + def create_pitch_file_path(file_name: str, cache_path: str) -> str: + return os.path.join(cache_path, file_name + "_pitch.npy") @staticmethod def _compute_and_save_pitch(ap, wav_file, pitch_file=None): @@ -773,7 +823,7 @@ def compute_pitch_stats(pitch_vecs): mean, std = np.mean(nonzeros), np.std(nonzeros) return mean, std - def load_stats(self, cache_path): + def load_stats(self, cache_path) -> None: stats_path = os.path.join(cache_path, "pitch_stats.npy") stats = np.load(stats_path, allow_pickle=True).item() self.mean = stats["mean"].astype(np.float32) @@ -794,9 +844,7 @@ def denormalize(self, pitch): return pitch def compute_or_load(self, wav_file, audio_unique_name): - """ - compute pitch and return a numpy array of pitch values - """ + """Compute pitch and return a numpy array of pitch values.""" pitch_file = self.create_pitch_file_path(audio_unique_name, self.cache_path) if not os.path.exists(pitch_file): pitch = self._compute_and_save_pitch(self.ap, wav_file, pitch_file) @@ -816,18 +864,18 @@ def collate_fn(self, batch): def print_logs(self, level: int = 0) -> None: indent = "\t" * level - print("\n") - print(f"{indent}> F0Dataset ") - print(f"{indent}| > Number of instances : {len(self.samples)}") + logger.info("%sF0Dataset", indent) + logger.info("%s| Number of instances : %d", indent, len(self.samples)) class EnergyDataset: - """Energy Dataset for computing Energy from wav files in CPU + """Energy Dataset for computing Energy from wav files in CPU. Pre-compute Energy values for all the samples at initialization if `cache_path` is not None or already present. It also computes the mean and std of Energy values if `normalize_Energy` is True. Args: + ---- samples (Union[List[List], List[Dict]]): List of samples. Each sample is a list or a dict. @@ -843,20 +891,19 @@ class EnergyDataset: normalize_Energy (bool): Whether to normalize Energy values by mean and std. Defaults to True. + """ def __init__( self, - samples: Union[List[List], List[Dict]], + samples: Union[list[list], list[dict]], ap: "AudioProcessor", - verbose=False, - cache_path: str = None, + cache_path: Optional[str] = None, precompute_num_workers=0, normalize_energy=True, - ): + ) -> None: self.samples = samples self.ap = ap - self.verbose = verbose self.cache_path = cache_path self.normalize_energy = normalize_energy self.pad_id = 0.0 @@ -876,18 +923,22 @@ def __getitem__(self, idx): energy = self.normalize(energy) return {"audio_unique_name": item["audio_unique_name"], "energy": energy} - def __len__(self): + def __len__(self) -> int: return len(self.samples) - def precompute(self, num_workers=0): - print("[*] Pre-computing energys...") + def precompute(self, num_workers=0) -> None: + logger.info("Pre-computing energys...") with tqdm.tqdm(total=len(self)) as pbar: batch_size = num_workers if num_workers > 0 else 1 # we do not normalize at preproessing normalize_energy = self.normalize_energy self.normalize_energy = False dataloder = torch.utils.data.DataLoader( - batch_size=batch_size, dataset=self, shuffle=False, num_workers=num_workers, collate_fn=self.collate_fn + batch_size=batch_size, + dataset=self, + shuffle=False, + num_workers=num_workers, + collate_fn=self.collate_fn, ) computed_data = [] for batch in dataloder: @@ -908,8 +959,7 @@ def get_pad_id(self): @staticmethod def create_energy_file_path(wav_file, cache_path): file_name = os.path.splitext(os.path.basename(wav_file))[0] - energy_file = os.path.join(cache_path, file_name + "_energy.npy") - return energy_file + return os.path.join(cache_path, file_name + "_energy.npy") @staticmethod def _compute_and_save_energy(ap, wav_file, energy_file=None): @@ -925,7 +975,7 @@ def compute_energy_stats(energy_vecs): mean, std = np.mean(nonzeros), np.std(nonzeros) return mean, std - def load_stats(self, cache_path): + def load_stats(self, cache_path) -> None: stats_path = os.path.join(cache_path, "energy_stats.npy") stats = np.load(stats_path, allow_pickle=True).item() self.mean = stats["mean"].astype(np.float32) @@ -946,9 +996,7 @@ def denormalize(self, energy): return energy def compute_or_load(self, wav_file, audio_unique_name): - """ - compute energy and return a numpy array of energy values - """ + """Compute energy and return a numpy array of energy values.""" energy_file = self.create_energy_file_path(audio_unique_name, self.cache_path) if not os.path.exists(energy_file): energy = self._compute_and_save_energy(self.ap, wav_file, energy_file) @@ -968,6 +1016,5 @@ def collate_fn(self, batch): def print_logs(self, level: int = 0) -> None: indent = "\t" * level - print("\n") - print(f"{indent}> energyDataset ") - print(f"{indent}| > Number of instances : {len(self.samples)}") + logger.info("%senergyDataset") + logger.info("%s| Number of instances : %d", indent, len(self.samples)) diff --git a/TTS/tts/datasets/formatters.py b/TTS/tts/datasets/formatters.py index 053444b0c1..ff1a76e2c9 100644 --- a/TTS/tts/datasets/formatters.py +++ b/TTS/tts/datasets/formatters.py @@ -1,3 +1,5 @@ +import csv +import logging import os import re import xml.etree.ElementTree as ET @@ -5,9 +7,10 @@ from pathlib import Path from typing import List -import pandas as pd from tqdm import tqdm +logger = logging.getLogger(__name__) + ######################## # DATASETS ######################## @@ -23,32 +26,34 @@ def cml_tts(root_path, meta_file, ignored_speakers=None): num_cols = len(lines[0].split("|")) # take the first row as reference for idx, line in enumerate(lines[1:]): if len(line.split("|")) != num_cols: - print(f" > Missing column in line {idx + 1} -> {line.strip()}") + logger.warning("Missing column in line %d -> %s", idx + 1, line.strip()) # load metadata - metadata = pd.read_csv(os.path.join(root_path, meta_file), sep="|") - assert all(x in metadata.columns for x in ["wav_filename", "transcript"]) - client_id = None if "client_id" in metadata.columns else "default" - emotion_name = None if "emotion_name" in metadata.columns else "neutral" + with open(Path(root_path) / meta_file, newline="", encoding="utf-8") as f: + reader = csv.DictReader(f, delimiter="|") + metadata = list(reader) + assert all(x in metadata[0] for x in ["wav_filename", "transcript"]) + client_id = None if "client_id" in metadata[0] else "default" + emotion_name = None if "emotion_name" in metadata[0] else "neutral" items = [] not_found_counter = 0 - for row in metadata.itertuples(): - if client_id is None and ignored_speakers is not None and row.client_id in ignored_speakers: + for row in metadata: + if client_id is None and ignored_speakers is not None and row["client_id"] in ignored_speakers: continue - audio_path = os.path.join(root_path, row.wav_filename) + audio_path = os.path.join(root_path, row["wav_filename"]) if not os.path.exists(audio_path): not_found_counter += 1 continue items.append( { - "text": row.transcript, + "text": row["transcript"], "audio_file": audio_path, - "speaker_name": client_id if client_id is not None else row.client_id, - "emotion_name": emotion_name if emotion_name is not None else row.emotion_name, + "speaker_name": client_id if client_id is not None else row["client_id"], + "emotion_name": emotion_name if emotion_name is not None else row["emotion_name"], "root_path": root_path, } ) if not_found_counter > 0: - print(f" | > [!] {not_found_counter} files not found") + logger.warning("%d files not found", not_found_counter) return items @@ -61,32 +66,34 @@ def coqui(root_path, meta_file, ignored_speakers=None): num_cols = len(lines[0].split("|")) # take the first row as reference for idx, line in enumerate(lines[1:]): if len(line.split("|")) != num_cols: - print(f" > Missing column in line {idx + 1} -> {line.strip()}") + logger.warning("Missing column in line %d -> %s", idx + 1, line.strip()) # load metadata - metadata = pd.read_csv(os.path.join(root_path, meta_file), sep="|") - assert all(x in metadata.columns for x in ["audio_file", "text"]) - speaker_name = None if "speaker_name" in metadata.columns else "coqui" - emotion_name = None if "emotion_name" in metadata.columns else "neutral" + with open(Path(root_path) / meta_file, newline="", encoding="utf-8") as f: + reader = csv.DictReader(f, delimiter="|") + metadata = list(reader) + assert all(x in metadata[0] for x in ["audio_file", "text"]) + speaker_name = None if "speaker_name" in metadata[0] else "coqui" + emotion_name = None if "emotion_name" in metadata[0] else "neutral" items = [] not_found_counter = 0 - for row in metadata.itertuples(): - if speaker_name is None and ignored_speakers is not None and row.speaker_name in ignored_speakers: + for row in metadata: + if speaker_name is None and ignored_speakers is not None and row["speaker_name"] in ignored_speakers: continue - audio_path = os.path.join(root_path, row.audio_file) + audio_path = os.path.join(root_path, row["audio_file"]) if not os.path.exists(audio_path): not_found_counter += 1 continue items.append( { - "text": row.text, + "text": row["text"], "audio_file": audio_path, - "speaker_name": speaker_name if speaker_name is not None else row.speaker_name, - "emotion_name": emotion_name if emotion_name is not None else row.emotion_name, + "speaker_name": speaker_name if speaker_name is not None else row["speaker_name"], + "emotion_name": emotion_name if emotion_name is not None else row["emotion_name"], "root_path": root_path, } ) if not_found_counter > 0: - print(f" | > [!] {not_found_counter} files not found") + logger.warning("%d files not found", not_found_counter) return items @@ -169,7 +176,7 @@ def mailabs(root_path, meta_files=None, ignored_speakers=None): if isinstance(ignored_speakers, list): if speaker_name in ignored_speakers: continue - print(" | > {}".format(csv_file)) + logger.info(csv_file) with open(txt_file, "r", encoding="utf-8") as ttf: for line in ttf: cols = line.split("|") @@ -184,7 +191,7 @@ def mailabs(root_path, meta_files=None, ignored_speakers=None): ) else: # M-AI-Labs have some missing samples, so just print the warning - print("> File %s does not exist!" % (wav_file)) + logger.warning("File %s does not exist!", wav_file) return items @@ -249,7 +256,7 @@ def sam_accenture(root_path, meta_file, **kwargs): # pylint: disable=unused-arg text = item.text wav_file = os.path.join(root_path, "vo_voice_quality_transformation", item.get("id") + ".wav") if not os.path.exists(wav_file): - print(f" [!] {wav_file} in metafile does not exist. Skipping...") + logger.warning("%s in metafile does not exist. Skipping...", wav_file) continue items.append({"text": text, "audio_file": wav_file, "speaker_name": speaker_name, "root_path": root_path}) return items @@ -370,7 +377,7 @@ def custom_turkish(root_path, meta_file, **kwargs): # pylint: disable=unused-ar continue text = cols[1].strip() items.append({"text": text, "audio_file": wav_file, "speaker_name": speaker_name, "root_path": root_path}) - print(f" [!] {len(skipped_files)} files skipped. They don't exist...") + logger.warning("%d files skipped. They don't exist...") return items @@ -438,7 +445,7 @@ def vctk(root_path, meta_files=None, wavs_path="wav48_silence_trimmed", mic="mic {"text": text, "audio_file": wav_file, "speaker_name": "VCTK_" + speaker_id, "root_path": root_path} ) else: - print(f" [!] wav files don't exist - {wav_file}") + logger.warning("Wav file doesn't exist - %s", wav_file) return items diff --git a/TTS/tts/layers/bark/hubert/hubert_manager.py b/TTS/tts/layers/bark/hubert/hubert_manager.py index 4bc1992941..fd936a9157 100644 --- a/TTS/tts/layers/bark/hubert/hubert_manager.py +++ b/TTS/tts/layers/bark/hubert/hubert_manager.py @@ -1,11 +1,14 @@ # From https://github.com/gitmylo/bark-voice-cloning-HuBERT-quantizer +import logging import os.path import shutil import urllib.request import huggingface_hub +logger = logging.getLogger(__name__) + class HubertManager: @staticmethod @@ -13,9 +16,9 @@ def make_sure_hubert_installed( download_url: str = "https://dl.fbaipublicfiles.com/hubert/hubert_base_ls960.pt", model_path: str = "" ): if not os.path.isfile(model_path): - print("Downloading HuBERT base model") + logger.info("Downloading HuBERT base model") urllib.request.urlretrieve(download_url, model_path) - print("Downloaded HuBERT") + logger.info("Downloaded HuBERT") return model_path return None @@ -27,9 +30,9 @@ def make_sure_tokenizer_installed( ): model_dir = os.path.dirname(model_path) if not os.path.isfile(model_path): - print("Downloading HuBERT custom tokenizer") + logger.info("Downloading HuBERT custom tokenizer") huggingface_hub.hf_hub_download(repo, model, local_dir=model_dir, local_dir_use_symlinks=False) shutil.move(os.path.join(model_dir, model), model_path) - print("Downloaded tokenizer") + logger.info("Downloaded tokenizer") return model_path return None diff --git a/TTS/tts/layers/bark/hubert/kmeans_hubert.py b/TTS/tts/layers/bark/hubert/kmeans_hubert.py index a6a3b9aeb1..ade84794eb 100644 --- a/TTS/tts/layers/bark/hubert/kmeans_hubert.py +++ b/TTS/tts/layers/bark/hubert/kmeans_hubert.py @@ -7,8 +7,6 @@ # Modified code from https://github.com/lucidrains/audiolm-pytorch/blob/main/audiolm_pytorch/hubert_kmeans.py -import logging -from pathlib import Path import torch from einops import pack, unpack @@ -16,6 +14,8 @@ from torchaudio.functional import resample from transformers import HubertModel +from TTS.utils.generic_utils import exists + def round_down_nearest_multiple(num, divisor): return num // divisor * divisor @@ -28,21 +28,13 @@ def curtail_to_multiple(t, mult, from_left=False): return t[..., seq_slice] -def exists(val): - return val is not None - - -def default(val, d): - return val if exists(val) else d - - class CustomHubert(nn.Module): """ checkpoint and kmeans can be downloaded at https://github.com/facebookresearch/fairseq/tree/main/examples/hubert or you can train your own """ - def __init__(self, checkpoint_path, target_sample_hz=16000, seq_len_multiple_of=None, output_layer=9, device=None): + def __init__(self, target_sample_hz=16000, seq_len_multiple_of=None, output_layer=9, device=None): super().__init__() self.target_sample_hz = target_sample_hz self.seq_len_multiple_of = seq_len_multiple_of diff --git a/TTS/tts/layers/bark/hubert/tokenizer.py b/TTS/tts/layers/bark/hubert/tokenizer.py index 3070241f1c..cd9579799a 100644 --- a/TTS/tts/layers/bark/hubert/tokenizer.py +++ b/TTS/tts/layers/bark/hubert/tokenizer.py @@ -5,6 +5,7 @@ """ import json +import logging import os.path from zipfile import ZipFile @@ -12,6 +13,8 @@ import torch from torch import nn, optim +logger = logging.getLogger(__name__) + class HubertTokenizer(nn.Module): def __init__(self, hidden_size=1024, input_size=768, output_size=10000, version=0): @@ -85,7 +88,7 @@ def train_step(self, x_train, y_train, log_loss=False): # Print loss if log_loss: - print("Loss", loss.item()) + logger.info("Loss %.3f", loss.item()) # Backward pass loss.backward() @@ -157,10 +160,10 @@ def auto_train(data_path, save_path="model.pth", load_model: str = None, save_ep data_x, data_y = [], [] if load_model and os.path.isfile(load_model): - print("Loading model from", load_model) + logger.info("Loading model from %s", load_model) model_training = HubertTokenizer.load_from_checkpoint(load_model, "cuda") else: - print("Creating new model.") + logger.info("Creating new model.") model_training = HubertTokenizer(version=1).to("cuda") # Settings for the model to run without lstm save_path = os.path.join(data_path, save_path) base_save_path = ".".join(save_path.split(".")[:-1]) @@ -191,5 +194,5 @@ def auto_train(data_path, save_path="model.pth", load_model: str = None, save_ep save_p_2 = f"{base_save_path}_epoch_{epoch}.pth" model_training.save(save_p) model_training.save(save_p_2) - print(f"Epoch {epoch} completed") + logger.info("Epoch %d completed", epoch) epoch += 1 diff --git a/TTS/tts/layers/bark/inference_funcs.py b/TTS/tts/layers/bark/inference_funcs.py index f3d3fee937..65c7800dcf 100644 --- a/TTS/tts/layers/bark/inference_funcs.py +++ b/TTS/tts/layers/bark/inference_funcs.py @@ -2,10 +2,11 @@ import os import re from glob import glob -from typing import Dict, List +from typing import Dict, List, Optional, Tuple import librosa import numpy as np +import numpy.typing as npt import torch import torchaudio import tqdm @@ -48,7 +49,7 @@ def get_voices(extra_voice_dirs: List[str] = []): # pylint: disable=dangerous-d return voices -def load_npz(npz_file): +def load_npz(npz_file: str) -> Tuple[npt.NDArray[np.int64], npt.NDArray[np.int64], npt.NDArray[np.int64]]: x_history = np.load(npz_file) semantic = x_history["semantic_prompt"] coarse = x_history["coarse_prompt"] @@ -56,7 +57,11 @@ def load_npz(npz_file): return semantic, coarse, fine -def load_voice(model, voice: str, extra_voice_dirs: List[str] = []): # pylint: disable=dangerous-default-value +def load_voice( + model, voice: str, extra_voice_dirs: List[str] = [] +) -> Tuple[ + Optional[npt.NDArray[np.int64]], Optional[npt.NDArray[np.int64]], Optional[npt.NDArray[np.int64]] +]: # pylint: disable=dangerous-default-value if voice == "random": return None, None, None @@ -107,11 +112,10 @@ def generate_voice( model, output_path, ): - """Generate a new voice from a given audio and text prompt. + """Generate a new voice from a given audio. Args: audio (np.ndarray): The audio to use as a base for the new voice. - text (str): Transcription of the audio you are clonning. model (BarkModel): The BarkModel to use for generating the new voice. output_path (str): The path to save the generated voice to. """ @@ -130,10 +134,9 @@ def generate_voice( # generate semantic tokens # Load the HuBERT model hubert_manager = HubertManager() - # hubert_manager.make_sure_hubert_installed(model_path=model.config.LOCAL_MODEL_PATHS["hubert"]) hubert_manager.make_sure_tokenizer_installed(model_path=model.config.LOCAL_MODEL_PATHS["hubert_tokenizer"]) - hubert_model = CustomHubert(checkpoint_path=model.config.LOCAL_MODEL_PATHS["hubert"]).to(model.device) + hubert_model = CustomHubert().to(model.device) # Load the CustomTokenizer model tokenizer = HubertTokenizer.load_from_checkpoint( diff --git a/TTS/tts/layers/bark/load_model.py b/TTS/tts/layers/bark/load_model.py index ce6b757f05..6b7caab916 100644 --- a/TTS/tts/layers/bark/load_model.py +++ b/TTS/tts/layers/bark/load_model.py @@ -10,14 +10,10 @@ from TTS.tts.layers.bark.model import GPT, GPTConfig from TTS.tts.layers.bark.model_fine import FineGPT, FineGPTConfig +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 -if ( - torch.cuda.is_available() - and hasattr(torch.cuda, "amp") - and hasattr(torch.cuda.amp, "autocast") - and torch.cuda.is_bf16_supported() -): - autocast = functools.partial(torch.cuda.amp.autocast, dtype=torch.bfloat16) +if torch.cuda.is_available() and torch.cuda.is_bf16_supported(): + autocast = functools.partial(torch.autocast, device_type="cuda", dtype=torch.bfloat16) else: @contextlib.contextmanager @@ -118,7 +114,7 @@ def load_model(ckpt_path, device, config, model_type="text"): logger.info(f"{model_type} model not found, downloading...") _download(config.REMOTE_MODEL_PATHS[model_type]["path"], ckpt_path, config.CACHE_DIR) - checkpoint = torch.load(ckpt_path, map_location=device) + checkpoint = torch.load(ckpt_path, map_location=device, weights_only=is_pytorch_at_least_2_4()) # this is a hack model_args = checkpoint["model_args"] if "input_vocab_size" not in model_args: diff --git a/TTS/tts/layers/bark/model.py b/TTS/tts/layers/bark/model.py index c84022bd08..54a9cecec0 100644 --- a/TTS/tts/layers/bark/model.py +++ b/TTS/tts/layers/bark/model.py @@ -2,6 +2,7 @@ Much of this code is adapted from Andrej Karpathy's NanoGPT (https://github.com/karpathy/nanoGPT) """ + import math from dataclasses import dataclass @@ -11,18 +12,6 @@ from torch.nn import functional as F -class LayerNorm(nn.Module): - """LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False""" - - def __init__(self, ndim, bias): - super().__init__() - self.weight = nn.Parameter(torch.ones(ndim)) - self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None - - def forward(self, x): - return F.layer_norm(x, self.weight.shape, self.weight, self.bias, 1e-5) - - class CausalSelfAttention(nn.Module): def __init__(self, config): super().__init__() @@ -118,9 +107,9 @@ def forward(self, x): class Block(nn.Module): def __init__(self, config, layer_idx): super().__init__() - self.ln_1 = LayerNorm(config.n_embd, bias=config.bias) + self.ln_1 = nn.LayerNorm(config.n_embd, bias=config.bias) self.attn = CausalSelfAttention(config) - self.ln_2 = LayerNorm(config.n_embd, bias=config.bias) + self.ln_2 = nn.LayerNorm(config.n_embd, bias=config.bias) self.mlp = MLP(config) self.layer_idx = layer_idx @@ -157,7 +146,7 @@ def __init__(self, config): wpe=nn.Embedding(config.block_size, config.n_embd), drop=nn.Dropout(config.dropout), h=nn.ModuleList([Block(config, idx) for idx in range(config.n_layer)]), - ln_f=LayerNorm(config.n_embd, bias=config.bias), + ln_f=nn.LayerNorm(config.n_embd, bias=config.bias), ) ) self.lm_head = nn.Linear(config.n_embd, config.output_vocab_size, bias=False) diff --git a/TTS/tts/layers/bark/model_fine.py b/TTS/tts/layers/bark/model_fine.py index 09e5f4765d..29126b41ab 100644 --- a/TTS/tts/layers/bark/model_fine.py +++ b/TTS/tts/layers/bark/model_fine.py @@ -2,6 +2,7 @@ Much of this code is adapted from Andrej Karpathy's NanoGPT (https://github.com/karpathy/nanoGPT) """ + import math from dataclasses import dataclass diff --git a/TTS/tts/layers/delightful_tts/acoustic_model.py b/TTS/tts/layers/delightful_tts/acoustic_model.py index c906b882e5..981d6cdb1f 100644 --- a/TTS/tts/layers/delightful_tts/acoustic_model.py +++ b/TTS/tts/layers/delightful_tts/acoustic_model.py @@ -1,16 +1,17 @@ ### credit: https://github.com/dunky11/voicesmith +import logging from typing import Callable, Dict, Tuple import torch import torch.nn.functional as F from coqpit import Coqpit +from monotonic_alignment_search import maximum_path from torch import nn from TTS.tts.layers.delightful_tts.conformer import Conformer from TTS.tts.layers.delightful_tts.encoders import ( PhonemeLevelProsodyEncoder, UtteranceLevelProsodyEncoder, - get_mask_from_lengths, ) from TTS.tts.layers.delightful_tts.energy_adaptor import EnergyAdaptor from TTS.tts.layers.delightful_tts.networks import EmbeddingPadded, positional_encoding @@ -18,7 +19,9 @@ from TTS.tts.layers.delightful_tts.pitch_adaptor import PitchAdaptor from TTS.tts.layers.delightful_tts.variance_predictor import VariancePredictor from TTS.tts.layers.generic.aligner import AlignmentNetwork -from TTS.tts.utils.helpers import generate_path, maximum_path, sequence_mask +from TTS.tts.utils.helpers import expand_encoder_outputs, generate_attention, sequence_mask + +logger = logging.getLogger(__name__) class AcousticModel(torch.nn.Module): @@ -217,7 +220,7 @@ def _set_speaker_input(self, aux_input: Dict): def _init_speaker_embedding(self): # pylint: disable=attribute-defined-outside-init if self.num_speakers > 0: - print(" > initialization of speaker-embedding layers.") + logger.info("Initialization of speaker-embedding layers.") self.embedded_speaker_dim = self.args.speaker_embedding_channels self.emb_g = nn.Embedding(self.num_speakers, self.embedded_speaker_dim) @@ -227,42 +230,6 @@ def _init_d_vector(self): raise ValueError("[!] Speaker embedding layer already initialized before d_vector settings.") self.embedded_speaker_dim = self.args.d_vector_dim - @staticmethod - def generate_attn(dr, x_mask, y_mask=None): - """Generate an attention mask from the linear scale durations. - - Args: - dr (Tensor): Linear scale durations. - x_mask (Tensor): Mask for the input (character) sequence. - y_mask (Tensor): Mask for the output (spectrogram) sequence. Compute it from the predicted durations - if None. Defaults to None. - - Shapes - - dr: :math:`(B, T_{en})` - - x_mask: :math:`(B, T_{en})` - - y_mask: :math:`(B, T_{de})` - """ - # compute decode mask from the durations - if y_mask is None: - y_lengths = dr.sum(1).long() - y_lengths[y_lengths < 1] = 1 - y_mask = torch.unsqueeze(sequence_mask(y_lengths, None), 1).to(dr.dtype) - attn_mask = torch.unsqueeze(x_mask, -1) * torch.unsqueeze(y_mask, 2) - attn = generate_path(dr, attn_mask.squeeze(1)).to(dr.dtype) - return attn - - def _expand_encoder_with_durations( - self, - o_en: torch.FloatTensor, - dr: torch.IntTensor, - x_mask: torch.IntTensor, - y_lengths: torch.IntTensor, - ): - y_mask = torch.unsqueeze(sequence_mask(y_lengths, None), 1).to(o_en.dtype) - attn = self.generate_attn(dr, x_mask, y_mask) - o_en_ex = torch.einsum("kmn, kjm -> kjn", [attn.float(), o_en]) - return y_mask, o_en_ex, attn.transpose(1, 2) - def _forward_aligner( self, x: torch.FloatTensor, @@ -336,8 +303,8 @@ def forward( {"d_vectors": d_vectors, "speaker_ids": speaker_idx} ) # pylint: disable=unused-variable - src_mask = get_mask_from_lengths(src_lens) # [B, T_src] - mel_mask = get_mask_from_lengths(mel_lens) # [B, T_mel] + src_mask = ~sequence_mask(src_lens) # [B, T_src] + mel_mask = ~sequence_mask(mel_lens) # [B, T_mel] # Token embeddings token_embeddings = self.src_word_emb(tokens) # [B, T_src, C_hidden] @@ -362,7 +329,7 @@ def forward( pos_encoding = positional_encoding( self.emb_dim, - max(token_embeddings.shape[1], max(mel_lens)), + max(token_embeddings.shape[1], *mel_lens), device=token_embeddings.device, ) encoder_outputs = self.encoder( @@ -416,8 +383,8 @@ def forward( encoder_outputs = encoder_outputs.transpose(1, 2) + pitch_emb + energy_emb log_duration_prediction = self.duration_predictor(x=encoder_outputs_res.detach(), mask=src_mask) - mel_pred_mask, encoder_outputs_ex, alignments = self._expand_encoder_with_durations( - o_en=encoder_outputs, y_lengths=mel_lens, dr=dr, x_mask=~src_mask[:, None] + encoder_outputs_ex, alignments, mel_pred_mask = expand_encoder_outputs( + encoder_outputs, y_lengths=mel_lens, duration=dr, x_mask=~src_mask[:, None] ) x = self.decoder( @@ -431,7 +398,7 @@ def forward( dr = torch.log(dr + 1) dr_pred = torch.exp(log_duration_prediction) - 1 - alignments_dp = self.generate_attn(dr_pred, src_mask.unsqueeze(1), mel_pred_mask) # [B, T_max, T_max2'] + alignments_dp = generate_attention(dr_pred, src_mask.unsqueeze(1), mel_pred_mask) # [B, T_max, T_max2'] return { "model_outputs": x, @@ -444,7 +411,7 @@ def forward( "p_prosody_pred": p_prosody_pred, "p_prosody_ref": p_prosody_ref, "alignments_dp": alignments_dp, - "alignments": alignments, # [B, T_de, T_en] + "alignments": alignments.transpose(1, 2), # [B, T_de, T_en] "aligner_soft": aligner_soft, "aligner_mas": aligner_mas, "aligner_durations": aligner_durations, @@ -465,7 +432,7 @@ def inference( pitch_transform: Callable = None, energy_transform: Callable = None, ) -> torch.Tensor: - src_mask = get_mask_from_lengths(torch.tensor([tokens.shape[1]], dtype=torch.int64, device=tokens.device)) + src_mask = ~sequence_mask(torch.tensor([tokens.shape[1]], dtype=torch.int64, device=tokens.device)) src_lens = torch.tensor(tokens.shape[1:2]).to(tokens.device) # pylint: disable=unused-variable sid, g, lid, _ = self._set_cond_input( # pylint: disable=unused-variable {"d_vectors": d_vectors, "speaker_ids": speaker_idx} @@ -532,11 +499,11 @@ def inference( duration_pred = torch.round(duration_pred) # -> [B, T_src] mel_lens = duration_pred.sum(1) # -> [B,] - _, encoder_outputs_ex, alignments = self._expand_encoder_with_durations( - o_en=encoder_outputs, y_lengths=mel_lens, dr=duration_pred.squeeze(1), x_mask=~src_mask[:, None] + encoder_outputs_ex, alignments, _ = expand_encoder_outputs( + encoder_outputs, y_lengths=mel_lens, duration=duration_pred.squeeze(1), x_mask=~src_mask[:, None] ) - mel_mask = get_mask_from_lengths( + mel_mask = ~sequence_mask( torch.tensor([encoder_outputs_ex.shape[2]], dtype=torch.int64, device=encoder_outputs_ex.device) ) @@ -553,7 +520,7 @@ def inference( x = self.to_mel(x) outputs = { "model_outputs": x, - "alignments": alignments, + "alignments": alignments.transpose(1, 2), # "pitch": pitch_emb_pred, "durations": duration_pred, "pitch": pitch_pred, diff --git a/TTS/tts/layers/delightful_tts/conformer.py b/TTS/tts/layers/delightful_tts/conformer.py index b2175b3b96..227a871c69 100644 --- a/TTS/tts/layers/delightful_tts/conformer.py +++ b/TTS/tts/layers/delightful_tts/conformer.py @@ -1,20 +1,14 @@ ### credit: https://github.com/dunky11/voicesmith import math -from typing import Tuple import torch import torch.nn as nn # pylint: disable=consider-using-from-import import torch.nn.functional as F -from TTS.tts.layers.delightful_tts.conv_layers import Conv1dGLU, DepthWiseConv1d, PointwiseConv1d +from TTS.tts.layers.delightful_tts.conv_layers import Conv1dGLU, DepthWiseConv1d, PointwiseConv1d, calc_same_padding from TTS.tts.layers.delightful_tts.networks import GLUActivation -def calc_same_padding(kernel_size: int) -> Tuple[int, int]: - pad = kernel_size // 2 - return (pad, pad - (kernel_size + 1) % 2) - - class Conformer(nn.Module): def __init__( self, @@ -322,7 +316,7 @@ def forward( value: torch.Tensor, mask: torch.Tensor, encoding: torch.Tensor, - ) -> Tuple[torch.Tensor, torch.Tensor]: + ) -> tuple[torch.Tensor, torch.Tensor]: batch_size, seq_length, _ = key.size() # pylint: disable=unused-variable encoding = encoding[:, : key.shape[1]] encoding = encoding.repeat(batch_size, 1, 1) @@ -378,7 +372,7 @@ def forward( value: torch.Tensor, pos_embedding: torch.Tensor, mask: torch.Tensor, - ) -> Tuple[torch.Tensor, torch.Tensor]: + ) -> tuple[torch.Tensor, torch.Tensor]: batch_size = query.shape[0] query = self.query_proj(query).view(batch_size, -1, self.num_heads, self.d_head) key = self.key_proj(key).view(batch_size, -1, self.num_heads, self.d_head).permute(0, 2, 1, 3) @@ -411,40 +405,3 @@ def _relative_shift(self, pos_score: torch.Tensor) -> torch.Tensor: # pylint: d padded_pos_score = padded_pos_score.view(batch_size, num_heads, seq_length2 + 1, seq_length1) pos_score = padded_pos_score[:, :, 1:].view_as(pos_score) return pos_score - - -class MultiHeadAttention(nn.Module): - """ - input: - query --- [N, T_q, query_dim] - key --- [N, T_k, key_dim] - output: - out --- [N, T_q, num_units] - """ - - def __init__(self, query_dim: int, key_dim: int, num_units: int, num_heads: int): - super().__init__() - self.num_units = num_units - self.num_heads = num_heads - self.key_dim = key_dim - - self.W_query = nn.Linear(in_features=query_dim, out_features=num_units, bias=False) - self.W_key = nn.Linear(in_features=key_dim, out_features=num_units, bias=False) - self.W_value = nn.Linear(in_features=key_dim, out_features=num_units, bias=False) - - def forward(self, query: torch.Tensor, key: torch.Tensor) -> torch.Tensor: - querys = self.W_query(query) # [N, T_q, num_units] - keys = self.W_key(key) # [N, T_k, num_units] - values = self.W_value(key) - split_size = self.num_units // self.num_heads - querys = torch.stack(torch.split(querys, split_size, dim=2), dim=0) # [h, N, T_q, num_units/h] - keys = torch.stack(torch.split(keys, split_size, dim=2), dim=0) # [h, N, T_k, num_units/h] - values = torch.stack(torch.split(values, split_size, dim=2), dim=0) # [h, N, T_k, num_units/h] - # score = softmax(QK^T / (d_k ** 0.5)) - scores = torch.matmul(querys, keys.transpose(2, 3)) # [h, N, T_q, T_k] - scores = scores / (self.key_dim**0.5) - scores = F.softmax(scores, dim=3) - # out = score * V - out = torch.matmul(scores, values) # [h, N, T_q, num_units/h] - out = torch.cat(torch.split(out, 1, dim=0), dim=3).squeeze(0) # [N, T_q, num_units] - return out diff --git a/TTS/tts/layers/delightful_tts/conv_layers.py b/TTS/tts/layers/delightful_tts/conv_layers.py index fb9aa4495f..1d5139571e 100644 --- a/TTS/tts/layers/delightful_tts/conv_layers.py +++ b/TTS/tts/layers/delightful_tts/conv_layers.py @@ -3,9 +3,6 @@ import torch import torch.nn as nn # pylint: disable=consider-using-from-import import torch.nn.functional as F -from torch.nn.utils import parametrize - -from TTS.tts.layers.delightful_tts.kernel_predictor import KernelPredictor def calc_same_padding(kernel_size: int) -> Tuple[int, int]: @@ -530,142 +527,3 @@ def forward(self, x: torch.Tensor) -> torch.Tensor: x = self.addcoords(x) x = self.conv(x) return x - - -class LVCBlock(torch.nn.Module): - """the location-variable convolutions""" - - def __init__( # pylint: disable=dangerous-default-value - self, - in_channels, - cond_channels, - stride, - dilations=[1, 3, 9, 27], - lReLU_slope=0.2, - conv_kernel_size=3, - cond_hop_length=256, - kpnet_hidden_channels=64, - kpnet_conv_size=3, - kpnet_dropout=0.0, - ): - super().__init__() - - self.cond_hop_length = cond_hop_length - self.conv_layers = len(dilations) - self.conv_kernel_size = conv_kernel_size - - self.kernel_predictor = KernelPredictor( - cond_channels=cond_channels, - conv_in_channels=in_channels, - conv_out_channels=2 * in_channels, - conv_layers=len(dilations), - conv_kernel_size=conv_kernel_size, - kpnet_hidden_channels=kpnet_hidden_channels, - kpnet_conv_size=kpnet_conv_size, - kpnet_dropout=kpnet_dropout, - kpnet_nonlinear_activation_params={"negative_slope": lReLU_slope}, - ) - - self.convt_pre = nn.Sequential( - nn.LeakyReLU(lReLU_slope), - nn.utils.parametrizations.weight_norm( - nn.ConvTranspose1d( - in_channels, - in_channels, - 2 * stride, - stride=stride, - padding=stride // 2 + stride % 2, - output_padding=stride % 2, - ) - ), - ) - - self.conv_blocks = nn.ModuleList() - for dilation in dilations: - self.conv_blocks.append( - nn.Sequential( - nn.LeakyReLU(lReLU_slope), - nn.utils.parametrizations.weight_norm( - nn.Conv1d( - in_channels, - in_channels, - conv_kernel_size, - padding=dilation * (conv_kernel_size - 1) // 2, - dilation=dilation, - ) - ), - nn.LeakyReLU(lReLU_slope), - ) - ) - - def forward(self, x, c): - """forward propagation of the location-variable convolutions. - Args: - x (Tensor): the input sequence (batch, in_channels, in_length) - c (Tensor): the conditioning sequence (batch, cond_channels, cond_length) - - Returns: - Tensor: the output sequence (batch, in_channels, in_length) - """ - _, in_channels, _ = x.shape # (B, c_g, L') - - x = self.convt_pre(x) # (B, c_g, stride * L') - kernels, bias = self.kernel_predictor(c) - - for i, conv in enumerate(self.conv_blocks): - output = conv(x) # (B, c_g, stride * L') - - k = kernels[:, i, :, :, :, :] # (B, 2 * c_g, c_g, kernel_size, cond_length) - b = bias[:, i, :, :] # (B, 2 * c_g, cond_length) - - output = self.location_variable_convolution( - output, k, b, hop_size=self.cond_hop_length - ) # (B, 2 * c_g, stride * L'): LVC - x = x + torch.sigmoid(output[:, :in_channels, :]) * torch.tanh( - output[:, in_channels:, :] - ) # (B, c_g, stride * L'): GAU - - return x - - def location_variable_convolution(self, x, kernel, bias, dilation=1, hop_size=256): # pylint: disable=no-self-use - """perform location-variable convolution operation on the input sequence (x) using the local convolution kernl. - Time: 414 Îŧs Âą 309 ns per loop (mean Âą std. dev. of 7 runs, 1000 loops each), test on NVIDIA V100. - Args: - x (Tensor): the input sequence (batch, in_channels, in_length). - kernel (Tensor): the local convolution kernel (batch, in_channel, out_channels, kernel_size, kernel_length) - bias (Tensor): the bias for the local convolution (batch, out_channels, kernel_length) - dilation (int): the dilation of convolution. - hop_size (int): the hop_size of the conditioning sequence. - Returns: - (Tensor): the output sequence after performing local convolution. (batch, out_channels, in_length). - """ - batch, _, in_length = x.shape - batch, _, out_channels, kernel_size, kernel_length = kernel.shape - assert in_length == (kernel_length * hop_size), "length of (x, kernel) is not matched" - - padding = dilation * int((kernel_size - 1) / 2) - x = F.pad(x, (padding, padding), "constant", 0) # (batch, in_channels, in_length + 2*padding) - x = x.unfold(2, hop_size + 2 * padding, hop_size) # (batch, in_channels, kernel_length, hop_size + 2*padding) - - if hop_size < dilation: - x = F.pad(x, (0, dilation), "constant", 0) - x = x.unfold( - 3, dilation, dilation - ) # (batch, in_channels, kernel_length, (hop_size + 2*padding)/dilation, dilation) - x = x[:, :, :, :, :hop_size] - x = x.transpose(3, 4) # (batch, in_channels, kernel_length, dilation, (hop_size + 2*padding)/dilation) - x = x.unfold(4, kernel_size, 1) # (batch, in_channels, kernel_length, dilation, _, kernel_size) - - o = torch.einsum("bildsk,biokl->bolsd", x, kernel) - o = o.to(memory_format=torch.channels_last_3d) - bias = bias.unsqueeze(-1).unsqueeze(-1).to(memory_format=torch.channels_last_3d) - o = o + bias - o = o.contiguous().view(batch, out_channels, -1) - - return o - - def remove_weight_norm(self): - self.kernel_predictor.remove_weight_norm() - parametrize.remove_parametrizations(self.convt_pre[1], "weight") - for block in self.conv_blocks: - parametrize.remove_parametrizations(block[1], "weight") diff --git a/TTS/tts/layers/delightful_tts/encoders.py b/TTS/tts/layers/delightful_tts/encoders.py index 0878f0677a..bd0c319dc1 100644 --- a/TTS/tts/layers/delightful_tts/encoders.py +++ b/TTS/tts/layers/delightful_tts/encoders.py @@ -7,14 +7,7 @@ from TTS.tts.layers.delightful_tts.conformer import ConformerMultiHeadedSelfAttention from TTS.tts.layers.delightful_tts.conv_layers import CoordConv1d from TTS.tts.layers.delightful_tts.networks import STL - - -def get_mask_from_lengths(lengths: torch.Tensor) -> torch.Tensor: - batch_size = lengths.shape[0] - max_len = torch.max(lengths).item() - ids = torch.arange(0, max_len, device=lengths.device).unsqueeze(0).expand(batch_size, -1) - mask = ids >= lengths.unsqueeze(1).expand(-1, max_len) - return mask +from TTS.tts.utils.helpers import sequence_mask def stride_lens(lens: torch.Tensor, stride: int = 2) -> torch.Tensor: @@ -93,7 +86,7 @@ def forward(self, x: torch.Tensor, mel_lens: torch.Tensor) -> Tuple[torch.Tensor outputs --- [N, E//2] """ - mel_masks = get_mask_from_lengths(mel_lens).unsqueeze(1) + mel_masks = ~sequence_mask(mel_lens).unsqueeze(1) x = x.masked_fill(mel_masks, 0) for conv, norm in zip(self.convs, self.norms): x = conv(x) @@ -103,7 +96,7 @@ def forward(self, x: torch.Tensor, mel_lens: torch.Tensor) -> Tuple[torch.Tensor for _ in range(2): mel_lens = stride_lens(mel_lens) - mel_masks = get_mask_from_lengths(mel_lens) + mel_masks = ~sequence_mask(mel_lens) x = x.masked_fill(mel_masks.unsqueeze(1), 0) x = x.permute((0, 2, 1)) diff --git a/TTS/tts/layers/delightful_tts/kernel_predictor.py b/TTS/tts/layers/delightful_tts/kernel_predictor.py deleted file mode 100644 index 96c550b6c2..0000000000 --- a/TTS/tts/layers/delightful_tts/kernel_predictor.py +++ /dev/null @@ -1,128 +0,0 @@ -import torch.nn as nn # pylint: disable=consider-using-from-import -from torch.nn.utils import parametrize - - -class KernelPredictor(nn.Module): - """Kernel predictor for the location-variable convolutions - - Args: - cond_channels (int): number of channel for the conditioning sequence, - conv_in_channels (int): number of channel for the input sequence, - conv_out_channels (int): number of channel for the output sequence, - conv_layers (int): number of layers - - """ - - def __init__( # pylint: disable=dangerous-default-value - self, - cond_channels, - conv_in_channels, - conv_out_channels, - conv_layers, - conv_kernel_size=3, - kpnet_hidden_channels=64, - kpnet_conv_size=3, - kpnet_dropout=0.0, - kpnet_nonlinear_activation="LeakyReLU", - kpnet_nonlinear_activation_params={"negative_slope": 0.1}, - ): - super().__init__() - - self.conv_in_channels = conv_in_channels - self.conv_out_channels = conv_out_channels - self.conv_kernel_size = conv_kernel_size - self.conv_layers = conv_layers - - kpnet_kernel_channels = conv_in_channels * conv_out_channels * conv_kernel_size * conv_layers # l_w - kpnet_bias_channels = conv_out_channels * conv_layers # l_b - - self.input_conv = nn.Sequential( - nn.utils.parametrizations.weight_norm( - nn.Conv1d(cond_channels, kpnet_hidden_channels, 5, padding=2, bias=True) - ), - getattr(nn, kpnet_nonlinear_activation)(**kpnet_nonlinear_activation_params), - ) - - self.residual_convs = nn.ModuleList() - padding = (kpnet_conv_size - 1) // 2 - for _ in range(3): - self.residual_convs.append( - nn.Sequential( - nn.Dropout(kpnet_dropout), - nn.utils.parametrizations.weight_norm( - nn.Conv1d( - kpnet_hidden_channels, - kpnet_hidden_channels, - kpnet_conv_size, - padding=padding, - bias=True, - ) - ), - getattr(nn, kpnet_nonlinear_activation)(**kpnet_nonlinear_activation_params), - nn.utils.parametrizations.weight_norm( - nn.Conv1d( - kpnet_hidden_channels, - kpnet_hidden_channels, - kpnet_conv_size, - padding=padding, - bias=True, - ) - ), - getattr(nn, kpnet_nonlinear_activation)(**kpnet_nonlinear_activation_params), - ) - ) - self.kernel_conv = nn.utils.parametrizations.weight_norm( - nn.Conv1d( - kpnet_hidden_channels, - kpnet_kernel_channels, - kpnet_conv_size, - padding=padding, - bias=True, - ) - ) - self.bias_conv = nn.utils.parametrizations.weight_norm( - nn.Conv1d( - kpnet_hidden_channels, - kpnet_bias_channels, - kpnet_conv_size, - padding=padding, - bias=True, - ) - ) - - def forward(self, c): - """ - Args: - c (Tensor): the conditioning sequence (batch, cond_channels, cond_length) - """ - batch, _, cond_length = c.shape - c = self.input_conv(c) - for residual_conv in self.residual_convs: - residual_conv.to(c.device) - c = c + residual_conv(c) - k = self.kernel_conv(c) - b = self.bias_conv(c) - kernels = k.contiguous().view( - batch, - self.conv_layers, - self.conv_in_channels, - self.conv_out_channels, - self.conv_kernel_size, - cond_length, - ) - bias = b.contiguous().view( - batch, - self.conv_layers, - self.conv_out_channels, - cond_length, - ) - - return kernels, bias - - def remove_weight_norm(self): - parametrize.remove_parametrizations(self.input_conv[0], "weight") - parametrize.remove_parametrizations(self.kernel_conv, "weight") - parametrize.remove_parametrizations(self.bias_conv, "weight") - for block in self.residual_convs: - parametrize.remove_parametrizations(block[1], "weight") - parametrize.remove_parametrizations(block[3], "weight") diff --git a/TTS/tts/layers/glow_tts/glow.py b/TTS/tts/layers/glow_tts/glow.py index b02c311808..77a796473b 100644 --- a/TTS/tts/layers/glow_tts/glow.py +++ b/TTS/tts/layers/glow_tts/glow.py @@ -1,5 +1,4 @@ import torch -from packaging.version import Version from torch import nn from torch.nn import functional as F @@ -90,10 +89,7 @@ def __init__(self, channels, num_splits=4, no_jacobian=False, **kwargs): # pyli self.no_jacobian = no_jacobian self.weight_inv = None - if Version(torch.__version__) < Version("1.9"): - w_init = torch.qr(torch.FloatTensor(self.num_splits, self.num_splits).normal_())[0] - else: - w_init = torch.linalg.qr(torch.FloatTensor(self.num_splits, self.num_splits).normal_(), "complete")[0] + w_init = torch.linalg.qr(torch.FloatTensor(self.num_splits, self.num_splits).normal_(), "complete")[0] if torch.det(w_init) < 0: w_init[:, 0] = -1 * w_init[:, 0] diff --git a/TTS/tts/layers/glow_tts/transformer.py b/TTS/tts/layers/glow_tts/transformer.py index 02688d611f..c97d070a95 100644 --- a/TTS/tts/layers/glow_tts/transformer.py +++ b/TTS/tts/layers/glow_tts/transformer.py @@ -5,6 +5,7 @@ from torch.nn import functional as F from TTS.tts.layers.generic.normalization import LayerNorm, LayerNorm2 +from TTS.tts.utils.helpers import convert_pad_shape class RelativePositionMultiHeadAttention(nn.Module): @@ -300,7 +301,7 @@ def _causal_padding(self, x): pad_l = self.kernel_size - 1 pad_r = 0 padding = [[0, 0], [0, 0], [pad_l, pad_r]] - x = F.pad(x, self._pad_shape(padding)) + x = F.pad(x, convert_pad_shape(padding)) return x def _same_padding(self, x): @@ -309,15 +310,9 @@ def _same_padding(self, x): pad_l = (self.kernel_size - 1) // 2 pad_r = self.kernel_size // 2 padding = [[0, 0], [0, 0], [pad_l, pad_r]] - x = F.pad(x, self._pad_shape(padding)) + x = F.pad(x, convert_pad_shape(padding)) return x - @staticmethod - def _pad_shape(padding): - l = padding[::-1] - pad_shape = [item for sublist in l for item in sublist] - return pad_shape - class RelativePositionTransformer(nn.Module): """Transformer with Relative Potional Encoding. diff --git a/TTS/tts/layers/losses.py b/TTS/tts/layers/losses.py index de5f408c48..db62430c9d 100644 --- a/TTS/tts/layers/losses.py +++ b/TTS/tts/layers/losses.py @@ -1,3 +1,4 @@ +import logging import math import numpy as np @@ -10,6 +11,8 @@ from TTS.tts.utils.ssim import SSIMLoss as _SSIMLoss from TTS.utils.audio.torch_transforms import TorchSTFT +logger = logging.getLogger(__name__) + # pylint: disable=abstract-method # relates https://github.com/pytorch/pytorch/issues/42305 @@ -132,11 +135,11 @@ def forward(self, y_hat, y, length): ssim_loss = self.loss_func((y_norm * mask).unsqueeze(1), (y_hat_norm * mask).unsqueeze(1)) if ssim_loss.item() > 1.0: - print(f" > SSIM loss is out-of-range {ssim_loss.item()}, setting it 1.0") + logger.info("SSIM loss is out-of-range (%.2f), setting it to 1.0", ssim_loss.item()) ssim_loss = torch.tensor(1.0, device=ssim_loss.device) if ssim_loss.item() < 0.0: - print(f" > SSIM loss is out-of-range {ssim_loss.item()}, setting it 0.0") + logger.info("SSIM loss is out-of-range (%.2f), setting it to 0.0", ssim_loss.item()) ssim_loss = torch.tensor(0.0, device=ssim_loss.device) return ssim_loss @@ -252,7 +255,7 @@ def forward(self, att_ws, ilens, olens): @staticmethod def _make_ga_mask(ilen, olen, sigma): - grid_x, grid_y = torch.meshgrid(torch.arange(olen).to(olen), torch.arange(ilen).to(ilen)) + grid_x, grid_y = torch.meshgrid(torch.arange(olen).to(olen), torch.arange(ilen).to(ilen), indexing="ij") grid_x, grid_y = grid_x.float(), grid_y.float() return 1.0 - torch.exp(-((grid_y / ilen - grid_x / olen) ** 2) / (2 * (sigma**2))) @@ -306,6 +309,24 @@ def forward(self, attn_logprob, in_lens, out_lens): return total_loss +class NLLLoss(nn.Module): + """Negative log likelihood loss.""" + + def forward(self, log_prob: torch.Tensor) -> dict: # pylint: disable=no-self-use + """Compute the loss. + + Args: + logits (Tensor): [B, T, D] + + Returns: + Tensor: [1] + + """ + return_dict = {} + return_dict["loss"] = -log_prob.mean() + return return_dict + + ######################## # MODEL LOSS LAYERS ######################## @@ -616,6 +637,28 @@ def forward( return {"loss": loss, "loss_l1": spec_loss, "loss_ssim": ssim_loss, "loss_dur": dur_loss, "mdn_loss": mdn_loss} +def feature_loss(feats_real, feats_generated): + loss = 0 + for dr, dg in zip(feats_real, feats_generated): + for rl, gl in zip(dr, dg): + rl = rl.float().detach() + gl = gl.float() + loss += torch.mean(torch.abs(rl - gl)) + return loss * 2 + + +def generator_loss(scores_fake): + loss = 0 + gen_losses = [] + for dg in scores_fake: + dg = dg.float() + l = torch.mean((1 - dg) ** 2) + gen_losses.append(l) + loss += l + + return loss, gen_losses + + class VitsGeneratorLoss(nn.Module): def __init__(self, c: Coqpit): super().__init__() @@ -637,28 +680,6 @@ def __init__(self, c: Coqpit): do_amp_to_db=True, ) - @staticmethod - def feature_loss(feats_real, feats_generated): - loss = 0 - for dr, dg in zip(feats_real, feats_generated): - for rl, gl in zip(dr, dg): - rl = rl.float().detach() - gl = gl.float() - loss += torch.mean(torch.abs(rl - gl)) - return loss * 2 - - @staticmethod - def generator_loss(scores_fake): - loss = 0 - gen_losses = [] - for dg in scores_fake: - dg = dg.float() - l = torch.mean((1 - dg) ** 2) - gen_losses.append(l) - loss += l - - return loss, gen_losses - @staticmethod def kl_loss(z_p, logs_q, m_p, logs_p, z_mask): """ @@ -719,10 +740,8 @@ def forward( self.kl_loss(z_p=z_p, logs_q=logs_q, m_p=m_p, logs_p=logs_p, z_mask=z_mask.unsqueeze(1)) * self.kl_loss_alpha ) - loss_feat = ( - self.feature_loss(feats_real=feats_disc_real, feats_generated=feats_disc_fake) * self.feat_loss_alpha - ) - loss_gen = self.generator_loss(scores_fake=scores_disc_fake)[0] * self.gen_loss_alpha + loss_feat = feature_loss(feats_real=feats_disc_real, feats_generated=feats_disc_fake) * self.feat_loss_alpha + loss_gen = generator_loss(scores_fake=scores_disc_fake)[0] * self.gen_loss_alpha loss_mel = torch.nn.functional.l1_loss(mel_slice, mel_slice_hat) * self.mel_loss_alpha loss_duration = torch.sum(loss_duration.float()) * self.dur_loss_alpha loss = loss_kl + loss_feat + loss_mel + loss_gen + loss_duration @@ -776,6 +795,15 @@ def forward(self, scores_disc_real, scores_disc_fake): return return_dict +def _binary_alignment_loss(alignment_hard, alignment_soft): + """Binary loss that forces soft alignments to match the hard alignments. + + Explained in `https://arxiv.org/pdf/2108.10447.pdf`. + """ + log_sum = torch.log(torch.clamp(alignment_soft[alignment_hard == 1], min=1e-12)).sum() + return -log_sum / alignment_hard.sum() + + class ForwardTTSLoss(nn.Module): """Generic configurable ForwardTTS loss.""" @@ -817,14 +845,6 @@ def __init__(self, c): self.dur_loss_alpha = c.dur_loss_alpha self.binary_alignment_loss_alpha = c.binary_align_loss_alpha - @staticmethod - def _binary_alignment_loss(alignment_hard, alignment_soft): - """Binary loss that forces soft alignments to match the hard alignments as - explained in `https://arxiv.org/pdf/2108.10447.pdf`. - """ - log_sum = torch.log(torch.clamp(alignment_soft[alignment_hard == 1], min=1e-12)).sum() - return -log_sum / alignment_hard.sum() - def forward( self, decoder_output, @@ -876,7 +896,7 @@ def forward( return_dict["loss_aligner"] = self.aligner_loss_alpha * aligner_loss if self.binary_alignment_loss_alpha > 0 and alignment_hard is not None: - binary_alignment_loss = self._binary_alignment_loss(alignment_hard, alignment_soft) + binary_alignment_loss = _binary_alignment_loss(alignment_hard, alignment_soft) loss = loss + self.binary_alignment_loss_alpha * binary_alignment_loss if binary_loss_weight: return_dict["loss_binary_alignment"] = ( diff --git a/TTS/tts/layers/overflow/common_layers.py b/TTS/tts/layers/overflow/common_layers.py index b036dd1bda..9f77af293c 100644 --- a/TTS/tts/layers/overflow/common_layers.py +++ b/TTS/tts/layers/overflow/common_layers.py @@ -1,3 +1,4 @@ +import logging from typing import List, Tuple import torch @@ -8,6 +9,8 @@ from TTS.tts.layers.tacotron.common_layers import Linear from TTS.tts.layers.tacotron.tacotron2 import ConvBNBlock +logger = logging.getLogger(__name__) + class Encoder(nn.Module): r"""Neural HMM Encoder @@ -213,8 +216,8 @@ def _floor_std(self, std): original_tensor = std.clone().detach() std = torch.clamp(std, min=self.std_floor) if torch.any(original_tensor != std): - print( - "[*] Standard deviation was floored! The model is preventing overfitting, nothing serious to worry about" + logger.info( + "Standard deviation was floored! The model is preventing overfitting, nothing serious to worry about" ) return std diff --git a/TTS/tts/layers/overflow/neural_hmm.py b/TTS/tts/layers/overflow/neural_hmm.py index 0631ba98c0..a12becef03 100644 --- a/TTS/tts/layers/overflow/neural_hmm.py +++ b/TTS/tts/layers/overflow/neural_hmm.py @@ -128,7 +128,8 @@ def forward(self, inputs, inputs_len, mels, mel_lens): # Get mean, std and transition vector from decoder for this timestep # Note: Gradient checkpointing currently doesn't works with multiple gpus inside a loop if self.use_grad_checkpointing and self.training: - mean, std, transition_vector = checkpoint(self.output_net, h_memory, inputs) + # TODO: use_reentrant=False is recommended + mean, std, transition_vector = checkpoint(self.output_net, h_memory, inputs, use_reentrant=True) else: mean, std, transition_vector = self.output_net(h_memory, inputs) diff --git a/TTS/tts/layers/overflow/plotting_utils.py b/TTS/tts/layers/overflow/plotting_utils.py index a63aeb370a..d9d3e3d141 100644 --- a/TTS/tts/layers/overflow/plotting_utils.py +++ b/TTS/tts/layers/overflow/plotting_utils.py @@ -71,7 +71,7 @@ def plot_transition_probabilities_to_numpy(states, transition_probabilities, out ax.set_title("Transition probability of state") ax.set_xlabel("hidden state") ax.set_ylabel("probability") - ax.set_xticks([i for i in range(len(transition_probabilities))]) # pylint: disable=unnecessary-comprehension + ax.set_xticks(list(range(len(transition_probabilities)))) ax.set_xticklabels([int(x) for x in states], rotation=90) plt.tight_layout() if not output_fig: diff --git a/TTS/tts/layers/tacotron/capacitron_layers.py b/TTS/tts/layers/tacotron/capacitron_layers.py index 2181ffa7ec..817f42771b 100644 --- a/TTS/tts/layers/tacotron/capacitron_layers.py +++ b/TTS/tts/layers/tacotron/capacitron_layers.py @@ -3,6 +3,8 @@ from torch.distributions.multivariate_normal import MultivariateNormal as MVN from torch.nn import functional as F +from TTS.tts.layers.tacotron.common_layers import calculate_post_conv_height + class CapacitronVAE(nn.Module): """Effective Use of Variational Embedding Capacity for prosody transfer. @@ -97,7 +99,7 @@ def __init__(self, num_mel, out_dim): self.training = False self.bns = nn.ModuleList([nn.BatchNorm2d(num_features=filter_size) for filter_size in filters[1:]]) - post_conv_height = self.calculate_post_conv_height(num_mel, 3, 2, 2, num_layers) + post_conv_height = calculate_post_conv_height(num_mel, 3, 2, 2, num_layers) self.recurrence = nn.LSTM( input_size=filters[-1] * post_conv_height, hidden_size=out_dim, batch_first=True, bidirectional=False ) @@ -155,13 +157,6 @@ def forward(self, inputs, input_lengths): return last_output.to(inputs.device) # [B, 128] - @staticmethod - def calculate_post_conv_height(height, kernel_size, stride, pad, n_convs): - """Height of spec after n convolutions with fixed kernel/stride/pad.""" - for _ in range(n_convs): - height = (height - kernel_size + 2 * pad) // stride + 1 - return height - class TextSummary(nn.Module): def __init__(self, embedding_dim, encoder_output_dim): diff --git a/TTS/tts/layers/tacotron/common_layers.py b/TTS/tts/layers/tacotron/common_layers.py index f78ff1e75f..16e517fdca 100644 --- a/TTS/tts/layers/tacotron/common_layers.py +++ b/TTS/tts/layers/tacotron/common_layers.py @@ -3,6 +3,13 @@ from torch.nn import functional as F +def calculate_post_conv_height(height: int, kernel_size: int, stride: int, pad: int, n_convs: int) -> int: + """Height of spec after n convolutions with fixed kernel/stride/pad.""" + for _ in range(n_convs): + height = (height - kernel_size + 2 * pad) // stride + 1 + return height + + class Linear(nn.Module): """Linear layer with a specific initialization. diff --git a/TTS/tts/layers/tacotron/gst_layers.py b/TTS/tts/layers/tacotron/gst_layers.py index 05dba7084f..4a83fb1c83 100644 --- a/TTS/tts/layers/tacotron/gst_layers.py +++ b/TTS/tts/layers/tacotron/gst_layers.py @@ -2,6 +2,8 @@ import torch.nn.functional as F from torch import nn +from TTS.tts.layers.tacotron.common_layers import calculate_post_conv_height + class GST(nn.Module): """Global Style Token Module for factorizing prosody in speech. @@ -44,7 +46,7 @@ def __init__(self, num_mel, embedding_dim): self.convs = nn.ModuleList(convs) self.bns = nn.ModuleList([nn.BatchNorm2d(num_features=filter_size) for filter_size in filters[1:]]) - post_conv_height = self.calculate_post_conv_height(num_mel, 3, 2, 1, num_layers) + post_conv_height = calculate_post_conv_height(num_mel, 3, 2, 1, num_layers) self.recurrence = nn.GRU( input_size=filters[-1] * post_conv_height, hidden_size=embedding_dim // 2, batch_first=True ) @@ -71,13 +73,6 @@ def forward(self, inputs): return out.squeeze(0) - @staticmethod - def calculate_post_conv_height(height, kernel_size, stride, pad, n_convs): - """Height of spec after n convolutions with fixed kernel/stride/pad.""" - for _ in range(n_convs): - height = (height - kernel_size + 2 * pad) // stride + 1 - return height - class StyleTokenLayer(nn.Module): """NN Module attending to style tokens based on prosody encodings.""" @@ -117,7 +112,7 @@ class MultiHeadAttention(nn.Module): out --- [N, T_q, num_units] """ - def __init__(self, query_dim, key_dim, num_units, num_heads): + def __init__(self, query_dim: int, key_dim: int, num_units: int, num_heads: int): super().__init__() self.num_units = num_units self.num_heads = num_heads @@ -127,7 +122,7 @@ def __init__(self, query_dim, key_dim, num_units, num_heads): self.W_key = nn.Linear(in_features=key_dim, out_features=num_units, bias=False) self.W_value = nn.Linear(in_features=key_dim, out_features=num_units, bias=False) - def forward(self, query, key): + def forward(self, query: torch.Tensor, key: torch.Tensor) -> torch.Tensor: queries = self.W_query(query) # [N, T_q, num_units] keys = self.W_key(key) # [N, T_k, num_units] values = self.W_value(key) @@ -137,13 +132,11 @@ def forward(self, query, key): keys = torch.stack(torch.split(keys, split_size, dim=2), dim=0) # [h, N, T_k, num_units/h] values = torch.stack(torch.split(values, split_size, dim=2), dim=0) # [h, N, T_k, num_units/h] - # score = softmax(QK^T / (d_k**0.5)) + # score = softmax(QK^T / (d_k ** 0.5)) scores = torch.matmul(queries, keys.transpose(2, 3)) # [h, N, T_q, T_k] scores = scores / (self.key_dim**0.5) scores = F.softmax(scores, dim=3) # out = score * V out = torch.matmul(scores, values) # [h, N, T_q, num_units/h] - out = torch.cat(torch.split(out, 1, dim=0), dim=3).squeeze(0) # [N, T_q, num_units] - - return out + return torch.cat(torch.split(out, 1, dim=0), dim=3).squeeze(0) # [N, T_q, num_units] diff --git a/TTS/tts/layers/tacotron/tacotron.py b/TTS/tts/layers/tacotron/tacotron.py index 7a47c35ef6..32643dfcee 100644 --- a/TTS/tts/layers/tacotron/tacotron.py +++ b/TTS/tts/layers/tacotron/tacotron.py @@ -1,12 +1,16 @@ # coding: utf-8 # adapted from https://github.com/r9y9/tacotron_pytorch +import logging + import torch from torch import nn from .attentions import init_attn from .common_layers import Prenet +logger = logging.getLogger(__name__) + class BatchNormConv1d(nn.Module): r"""A wrapper for Conv1d with BatchNorm. It sets the activation @@ -480,7 +484,7 @@ def inference(self, inputs): if t > inputs.shape[1] / 4 and (stop_token > 0.6 or attention[:, -1].item() > 0.6): break if t > self.max_decoder_steps: - print(" | > Decoder stopped with 'max_decoder_steps") + logger.info("Decoder stopped with `max_decoder_steps` %d", self.max_decoder_steps) break return self._parse_outputs(outputs, attentions, stop_tokens) diff --git a/TTS/tts/layers/tacotron/tacotron2.py b/TTS/tts/layers/tacotron/tacotron2.py index c79b709972..727bf9ecfd 100644 --- a/TTS/tts/layers/tacotron/tacotron2.py +++ b/TTS/tts/layers/tacotron/tacotron2.py @@ -1,3 +1,5 @@ +import logging + import torch from torch import nn from torch.nn import functional as F @@ -5,6 +7,8 @@ from .attentions import init_attn from .common_layers import Linear, Prenet +logger = logging.getLogger(__name__) + # pylint: disable=no-value-for-parameter # pylint: disable=unexpected-keyword-arg @@ -356,7 +360,7 @@ def inference(self, inputs): if stop_token > self.stop_threshold and t > inputs.shape[0] // 2: break if len(outputs) == self.max_decoder_steps: - print(f" > Decoder stopped with `max_decoder_steps` {self.max_decoder_steps}") + logger.info("Decoder stopped with `max_decoder_steps` %d", self.max_decoder_steps) break memory = self._update_memory(decoder_output) @@ -389,7 +393,7 @@ def inference_truncated(self, inputs): if stop_token > 0.7: break if len(outputs) == self.max_decoder_steps: - print(" | > Decoder stopped with 'max_decoder_steps") + logger.info("Decoder stopped with `max_decoder_steps` %d", self.max_decoder_steps) break self.memory_truncated = decoder_output diff --git a/TTS/tts/layers/tortoise/arch_utils.py b/TTS/tts/layers/tortoise/arch_utils.py index dad1814369..1bbf676393 100644 --- a/TTS/tts/layers/tortoise/arch_utils.py +++ b/TTS/tts/layers/tortoise/arch_utils.py @@ -1,6 +1,5 @@ import functools import math -import os import fsspec import torch @@ -10,6 +9,7 @@ from transformers import LogitsWarper from TTS.tts.layers.tortoise.xtransformers import ContinuousTransformerWrapper, RelativePositionBias +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 def zero_module(module): @@ -70,11 +70,10 @@ def forward(self, qkv, mask=None, rel_pos=None): weight = rel_pos(weight.reshape(bs, self.n_heads, weight.shape[-2], weight.shape[-1])).reshape( bs * self.n_heads, weight.shape[-2], weight.shape[-1] ) - weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) if mask is not None: - # The proper way to do this is to mask before the softmax using -inf, but that doesn't work properly on CPUs. - mask = mask.repeat(self.n_heads, 1).unsqueeze(1) - weight = weight * mask + mask = mask.repeat(self.n_heads, 1, 1) + weight[mask.logical_not()] = -torch.inf + weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) a = torch.einsum("bts,bcs->bct", weight, v) return a.reshape(bs, -1, length) @@ -93,12 +92,12 @@ def __init__( channels, num_heads=1, num_head_channels=-1, - do_checkpoint=True, + *, relative_pos_embeddings=False, + tortoise_norm=False, ): super().__init__() self.channels = channels - self.do_checkpoint = do_checkpoint if num_head_channels == -1: self.num_heads = num_heads else: @@ -110,6 +109,7 @@ def __init__( self.qkv = nn.Conv1d(channels, channels * 3, 1) # split heads before split qkv self.attention = QKVAttentionLegacy(self.num_heads) + self.tortoise_norm = tortoise_norm self.proj_out = zero_module(nn.Conv1d(channels, channels, 1)) if relative_pos_embeddings: @@ -126,10 +126,13 @@ def __init__( def forward(self, x, mask=None): b, c, *spatial = x.shape x = x.reshape(b, c, -1) - qkv = self.qkv(self.norm(x)) + x_norm = self.norm(x) + qkv = self.qkv(x_norm) h = self.attention(qkv, mask, self.relative_pos_embeddings) h = self.proj_out(h) - return (x + h).reshape(b, c, *spatial) + if self.tortoise_norm: + return (x + h).reshape(b, c, *spatial) + return (x_norm + h).reshape(b, c, *spatial) class Upsample(nn.Module): @@ -185,115 +188,7 @@ def forward(self, x): return self.op(x) -class ResBlock(nn.Module): - def __init__( - self, - channels, - dropout, - out_channels=None, - use_conv=False, - use_scale_shift_norm=False, - up=False, - down=False, - kernel_size=3, - ): - super().__init__() - self.channels = channels - self.dropout = dropout - self.out_channels = out_channels or channels - self.use_conv = use_conv - self.use_scale_shift_norm = use_scale_shift_norm - padding = 1 if kernel_size == 3 else 2 - - self.in_layers = nn.Sequential( - normalization(channels), - nn.SiLU(), - nn.Conv1d(channels, self.out_channels, kernel_size, padding=padding), - ) - - self.updown = up or down - - if up: - self.h_upd = Upsample(channels, False) - self.x_upd = Upsample(channels, False) - elif down: - self.h_upd = Downsample(channels, False) - self.x_upd = Downsample(channels, False) - else: - self.h_upd = self.x_upd = nn.Identity() - - self.out_layers = nn.Sequential( - normalization(self.out_channels), - nn.SiLU(), - nn.Dropout(p=dropout), - zero_module(nn.Conv1d(self.out_channels, self.out_channels, kernel_size, padding=padding)), - ) - - if self.out_channels == channels: - self.skip_connection = nn.Identity() - elif use_conv: - self.skip_connection = nn.Conv1d(channels, self.out_channels, kernel_size, padding=padding) - else: - self.skip_connection = nn.Conv1d(channels, self.out_channels, 1) - - def forward(self, x): - if self.updown: - in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] - h = in_rest(x) - h = self.h_upd(h) - x = self.x_upd(x) - h = in_conv(h) - else: - h = self.in_layers(x) - h = self.out_layers(h) - return self.skip_connection(x) + h - - -class AudioMiniEncoder(nn.Module): - def __init__( - self, - spec_dim, - embedding_dim, - base_channels=128, - depth=2, - resnet_blocks=2, - attn_blocks=4, - num_attn_heads=4, - dropout=0, - downsample_factor=2, - kernel_size=3, - ): - super().__init__() - self.init = nn.Sequential(nn.Conv1d(spec_dim, base_channels, 3, padding=1)) - ch = base_channels - res = [] - for l in range(depth): - for r in range(resnet_blocks): - res.append(ResBlock(ch, dropout, kernel_size=kernel_size)) - res.append(Downsample(ch, use_conv=True, out_channels=ch * 2, factor=downsample_factor)) - ch *= 2 - self.res = nn.Sequential(*res) - self.final = nn.Sequential(normalization(ch), nn.SiLU(), nn.Conv1d(ch, embedding_dim, 1)) - attn = [] - for a in range(attn_blocks): - attn.append( - AttentionBlock( - embedding_dim, - num_attn_heads, - ) - ) - self.attn = nn.Sequential(*attn) - self.dim = embedding_dim - - def forward(self, x): - h = self.init(x) - h = self.res(h) - h = self.final(h) - h = self.attn(h) - return h[:, :, 0] - - -DEFAULT_MEL_NORM_FILE = "https://coqui.gateway.scarf.sh/v0.14.1_models/mel_norms.pth" +DEFAULT_MEL_NORM_FILE = "https://github.com/coqui-ai/TTS/releases/download/v0.14.1_models/mel_norms.pth" class TorchMelSpectrogram(nn.Module): @@ -333,7 +228,7 @@ def __init__( self.mel_norm_file = mel_norm_file if self.mel_norm_file is not None: with fsspec.open(self.mel_norm_file) as f: - self.mel_norms = torch.load(f) + self.mel_norms = torch.load(f, weights_only=is_pytorch_at_least_2_4()) else: self.mel_norms = None diff --git a/TTS/tts/layers/tortoise/audio_utils.py b/TTS/tts/layers/tortoise/audio_utils.py index 70711ed7a4..c67ee6c44b 100644 --- a/TTS/tts/layers/tortoise/audio_utils.py +++ b/TTS/tts/layers/tortoise/audio_utils.py @@ -1,3 +1,4 @@ +import logging import os from glob import glob from typing import Dict, List @@ -8,7 +9,10 @@ import torchaudio from scipy.io.wavfile import read -from TTS.utils.audio.torch_transforms import TorchSTFT +from TTS.utils.audio.torch_transforms import TorchSTFT, amp_to_db +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 + +logger = logging.getLogger(__name__) def load_wav_to_torch(full_path): @@ -28,7 +32,7 @@ def check_audio(audio, audiopath: str): # Check some assumptions about audio range. This should be automatically fixed in load_wav_to_torch, but might not be in some edge cases, where we should squawk. # '2' is arbitrarily chosen since it seems like audio will often "overdrive" the [-1,1] bounds. if torch.any(audio > 2) or not torch.any(audio < 0): - print(f"Error with {audiopath}. Max={audio.max()} min={audio.min()}") + logger.error("Error with %s. Max=%.2f min=%.2f", audiopath, audio.max(), audio.min()) audio.clip_(-1, 1) @@ -84,24 +88,6 @@ def normalize_tacotron_mel(mel): return 2 * ((mel - TACOTRON_MEL_MIN) / (TACOTRON_MEL_MAX - TACOTRON_MEL_MIN)) - 1 -def dynamic_range_compression(x, C=1, clip_val=1e-5): - """ - PARAMS - ------ - C: compression factor - """ - return torch.log(torch.clamp(x, min=clip_val) * C) - - -def dynamic_range_decompression(x, C=1): - """ - PARAMS - ------ - C: compression factor used to compress - """ - return torch.exp(x) / C - - def get_voices(extra_voice_dirs: List[str] = []): dirs = extra_voice_dirs voices: Dict[str, List[str]] = {} @@ -121,7 +107,7 @@ def load_voice(voice: str, extra_voice_dirs: List[str] = []): voices = get_voices(extra_voice_dirs) paths = voices[voice] if len(paths) == 1 and paths[0].endswith(".pth"): - return None, torch.load(paths[0]) + return None, torch.load(paths[0], weights_only=is_pytorch_at_least_2_4()) else: conds = [] for cond_path in paths: @@ -136,7 +122,7 @@ def load_voices(voices: List[str], extra_voice_dirs: List[str] = []): for voice in voices: if voice == "random": if len(voices) > 1: - print("Cannot combine a random voice with a non-random voice. Just using a random voice.") + logger.warning("Cannot combine a random voice with a non-random voice. Just using a random voice.") return None, None clip, latent = load_voice(voice, extra_voice_dirs) if latent is None: @@ -171,7 +157,7 @@ def wav_to_univnet_mel(wav, do_normalization=False, device="cuda"): ) stft = stft.to(device) mel = stft(wav) - mel = dynamic_range_compression(mel) + mel = amp_to_db(mel) if do_normalization: mel = normalize_tacotron_mel(mel) return mel diff --git a/TTS/tts/layers/tortoise/autoregressive.py b/TTS/tts/layers/tortoise/autoregressive.py index 14d881bc10..00c884e973 100644 --- a/TTS/tts/layers/tortoise/autoregressive.py +++ b/TTS/tts/layers/tortoise/autoregressive.py @@ -1,14 +1,23 @@ # AGPL: a notification must be added stating that changes have been made to that file. import functools +import random +from typing import Optional import torch import torch.nn as nn import torch.nn.functional as F +import transformers +from packaging.version import Version from transformers import GPT2Config, GPT2PreTrainedModel, LogitsProcessorList from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions from TTS.tts.layers.tortoise.arch_utils import AttentionBlock, TypicalLogitsWarper +if Version(transformers.__version__) >= Version("4.45"): + isin = transformers.pytorch_utils.isin_mps_friendly +else: + isin = torch.isin + def null_position_embeddings(range, dim): return torch.zeros((range.shape[0], range.shape[1], dim), device=range.device) @@ -115,7 +124,7 @@ def forward( else: emb = self.embeddings(input_ids) emb = emb + self.text_pos_embedding.get_fixed_embedding( - attention_mask.shape[1] - mel_len, attention_mask.device + attention_mask.shape[1] - (mel_len + 1), attention_mask.device ) transformer_outputs = self.transformer( @@ -167,44 +176,56 @@ def __init__( embedding_dim, attn_blocks=6, num_attn_heads=4, - do_checkpointing=False, - mean=False, + *, + tortoise_norm=False, ): super().__init__() attn = [] self.init = nn.Conv1d(spec_dim, embedding_dim, kernel_size=1) for a in range(attn_blocks): - attn.append(AttentionBlock(embedding_dim, num_attn_heads)) + attn.append(AttentionBlock(embedding_dim, num_attn_heads, tortoise_norm=tortoise_norm)) self.attn = nn.Sequential(*attn) self.dim = embedding_dim - self.do_checkpointing = do_checkpointing - self.mean = mean def forward(self, x): + """ + x: (b, 80, s) + """ h = self.init(x) h = self.attn(h) - if self.mean: - return h.mean(dim=2) - else: - return h[:, :, 0] + return h class LearnedPositionEmbeddings(nn.Module): - def __init__(self, seq_len, model_dim, init=0.02): + def __init__(self, seq_len, model_dim, init=0.02, relative=False): super().__init__() self.emb = nn.Embedding(seq_len, model_dim) # Initializing this way is standard for GPT-2 self.emb.weight.data.normal_(mean=0.0, std=init) + self.relative = relative + self.seq_len = seq_len def forward(self, x): sl = x.shape[1] - return self.emb(torch.arange(0, sl, device=x.device)) + if self.relative: + start = random.randint(sl, self.seq_len) - sl + return self.emb(torch.arange(start, start + sl, device=x.device)) + else: + return self.emb(torch.arange(0, sl, device=x.device)) def get_fixed_embedding(self, ind, dev): - return self.emb(torch.arange(0, ind, device=dev))[ind - 1 : ind] - - -def build_hf_gpt_transformer(layers, model_dim, heads, max_mel_seq_len, max_text_seq_len, checkpointing): + return self.emb(torch.tensor([ind], device=dev)).unsqueeze(0) + + +def build_hf_gpt_transformer( + layers: int, + model_dim: int, + heads: int, + max_mel_seq_len: int, + max_text_seq_len: int, + checkpointing: bool, + max_prompt_len: int = 0, +): """ GPT-2 implemented by the HuggingFace library. """ @@ -212,8 +233,8 @@ def build_hf_gpt_transformer(layers, model_dim, heads, max_mel_seq_len, max_text gpt_config = GPT2Config( vocab_size=256, # Unused. - n_positions=max_mel_seq_len + max_text_seq_len, - n_ctx=max_mel_seq_len + max_text_seq_len, + n_positions=max_mel_seq_len + max_text_seq_len + max_prompt_len, + n_ctx=max_mel_seq_len + max_text_seq_len + max_prompt_len, n_embd=model_dim, n_layer=layers, n_head=heads, @@ -226,13 +247,18 @@ def build_hf_gpt_transformer(layers, model_dim, heads, max_mel_seq_len, max_text gpt.wpe = functools.partial(null_position_embeddings, dim=model_dim) # Built-in token embeddings are unused. del gpt.wte - return ( - gpt, - LearnedPositionEmbeddings(max_mel_seq_len, model_dim), - LearnedPositionEmbeddings(max_text_seq_len, model_dim), - None, - None, + + mel_pos_emb = ( + LearnedPositionEmbeddings(max_mel_seq_len, model_dim) + if max_mel_seq_len != -1 + else functools.partial(null_position_embeddings, dim=model_dim) + ) + text_pos_emb = ( + LearnedPositionEmbeddings(max_text_seq_len, model_dim) + if max_mel_seq_len != -1 + else functools.partial(null_position_embeddings, dim=model_dim) ) + return gpt, mel_pos_emb, text_pos_emb, None, None class MelEncoder(nn.Module): @@ -326,12 +352,12 @@ def __init__( self.mel_layer_pos_embedding, self.text_layer_pos_embedding, ) = build_hf_gpt_transformer( - layers, - model_dim, - heads, - self.max_mel_tokens + 2 + self.max_conditioning_inputs, - self.max_text_tokens + 2, - checkpointing, + layers=layers, + model_dim=model_dim, + heads=heads, + max_mel_seq_len=self.max_mel_tokens + 2 + self.max_conditioning_inputs, + max_text_seq_len=self.max_text_tokens + 2, + checkpointing=checkpointing, ) if train_solo_embeddings: self.mel_solo_embedding = nn.Parameter(torch.randn(1, 1, model_dim) * 0.02, requires_grad=True) @@ -447,7 +473,7 @@ def get_conditioning(self, speech_conditioning_input): ) conds = [] for j in range(speech_conditioning_input.shape[1]): - conds.append(self.conditioning_encoder(speech_conditioning_input[:, j])) + conds.append(self.conditioning_encoder(speech_conditioning_input[:, j])[:, :, 0]) conds = torch.stack(conds, dim=1) conds = conds.mean(dim=1) return conds @@ -596,6 +622,8 @@ def inference_speech( max_length = ( trunc_index + self.max_mel_tokens - 1 if max_generate_length is None else trunc_index + max_generate_length ) + stop_token_tensor = torch.tensor(self.stop_mel_token, device=inputs.device, dtype=torch.long) + attention_mask = _prepare_attention_mask_for_generation(inputs, stop_token_tensor, stop_token_tensor) gen = self.inference_model.generate( inputs, bos_token_id=self.start_mel_token, @@ -604,11 +632,39 @@ def inference_speech( max_length=max_length, logits_processor=logits_processor, num_return_sequences=num_return_sequences, + attention_mask=attention_mask, **hf_generate_kwargs, ) return gen[:, trunc_index:] +def _prepare_attention_mask_for_generation( + inputs: torch.Tensor, + pad_token_id: Optional[torch.Tensor], + eos_token_id: Optional[torch.Tensor], +) -> torch.LongTensor: + # No information for attention mask inference -> return default attention mask + default_attention_mask = torch.ones(inputs.shape[:2], dtype=torch.long, device=inputs.device) + if pad_token_id is None: + return default_attention_mask + + is_input_ids = len(inputs.shape) == 2 and inputs.dtype in [torch.int, torch.long] + if not is_input_ids: + return default_attention_mask + + is_pad_token_in_inputs = (pad_token_id is not None) and (isin(elements=inputs, test_elements=pad_token_id).any()) + is_pad_token_not_equal_to_eos_token_id = (eos_token_id is None) or ~( + isin(elements=eos_token_id, test_elements=pad_token_id).any() + ) + can_infer_attention_mask = is_pad_token_in_inputs * is_pad_token_not_equal_to_eos_token_id + attention_mask_from_padding = inputs.ne(pad_token_id).long() + + attention_mask = ( + attention_mask_from_padding * can_infer_attention_mask + default_attention_mask * ~can_infer_attention_mask + ) + return attention_mask + + if __name__ == "__main__": gpt = UnifiedVoice( model_dim=256, diff --git a/TTS/tts/layers/tortoise/classifier.py b/TTS/tts/layers/tortoise/classifier.py index 8764bb070b..337323db67 100644 --- a/TTS/tts/layers/tortoise/classifier.py +++ b/TTS/tts/layers/tortoise/classifier.py @@ -16,7 +16,6 @@ def __init__( up=False, down=False, kernel_size=3, - do_checkpoint=True, ): super().__init__() self.channels = channels @@ -24,7 +23,6 @@ def __init__( self.out_channels = out_channels or channels self.use_conv = use_conv self.use_scale_shift_norm = use_scale_shift_norm - self.do_checkpoint = do_checkpoint padding = 1 if kernel_size == 3 else 2 self.in_layers = nn.Sequential( @@ -92,14 +90,14 @@ def __init__( self.layers = depth for l in range(depth): for r in range(resnet_blocks): - res.append(ResBlock(ch, dropout, do_checkpoint=False, kernel_size=kernel_size)) + res.append(ResBlock(ch, dropout, kernel_size=kernel_size)) res.append(Downsample(ch, use_conv=True, out_channels=ch * 2, factor=downsample_factor)) ch *= 2 self.res = nn.Sequential(*res) self.final = nn.Sequential(normalization(ch), nn.SiLU(), nn.Conv1d(ch, embedding_dim, 1)) attn = [] for a in range(attn_blocks): - attn.append(AttentionBlock(embedding_dim, num_attn_heads, do_checkpoint=False)) + attn.append(AttentionBlock(embedding_dim, num_attn_heads, tortoise_norm=True)) self.attn = nn.Sequential(*attn) self.dim = embedding_dim diff --git a/TTS/tts/layers/tortoise/clvp.py b/TTS/tts/layers/tortoise/clvp.py index 69b8c17c3f..44da1324e7 100644 --- a/TTS/tts/layers/tortoise/clvp.py +++ b/TTS/tts/layers/tortoise/clvp.py @@ -8,10 +8,6 @@ from TTS.tts.layers.tortoise.xtransformers import Encoder -def exists(val): - return val is not None - - def masked_mean(t, mask, dim=1): t = t.masked_fill(~mask[:, :, None], 0.0) return t.sum(dim=1) / mask.sum(dim=1)[..., None] @@ -126,7 +122,7 @@ def forward(self, text, speech_tokens, return_loss=False): text_latents = self.to_text_latent(text_latents) speech_latents = self.to_speech_latent(speech_latents) - text_latents, speech_latents = map(lambda t: F.normalize(t, p=2, dim=-1), (text_latents, speech_latents)) + text_latents, speech_latents = (F.normalize(t, p=2, dim=-1) for t in (text_latents, speech_latents)) temp = self.temperature.exp() diff --git a/TTS/tts/layers/tortoise/diffusion.py b/TTS/tts/layers/tortoise/diffusion.py index 7bea02ca08..2b29091b44 100644 --- a/TTS/tts/layers/tortoise/diffusion.py +++ b/TTS/tts/layers/tortoise/diffusion.py @@ -972,7 +972,7 @@ def autoregressive_training_losses( assert False # not currently supported for this type of diffusion. elif self.loss_type == LossType.MSE or self.loss_type == LossType.RESCALED_MSE: model_outputs = model(x_t, x_start, self._scale_timesteps(t), **model_kwargs) - terms.update({k: o for k, o in zip(model_output_keys, model_outputs)}) + terms.update(dict(zip(model_output_keys, model_outputs))) model_output = terms[gd_out_key] if self.model_var_type in [ ModelVarType.LEARNED, diff --git a/TTS/tts/layers/tortoise/diffusion_decoder.py b/TTS/tts/layers/tortoise/diffusion_decoder.py index 0d3cf7698a..cfdeaff8bb 100644 --- a/TTS/tts/layers/tortoise/diffusion_decoder.py +++ b/TTS/tts/layers/tortoise/diffusion_decoder.py @@ -5,7 +5,6 @@ import torch import torch.nn as nn import torch.nn.functional as F -from torch import autocast from TTS.tts.layers.tortoise.arch_utils import AttentionBlock, normalization @@ -131,7 +130,7 @@ def __init__(self, model_channels, dropout, num_heads): dims=1, use_scale_shift_norm=True, ) - self.attn = AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True) + self.attn = AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True, tortoise_norm=True) def forward(self, x, time_emb): y = self.resblk(x, time_emb) @@ -178,17 +177,17 @@ def __init__( # transformer network. self.code_embedding = nn.Embedding(in_tokens, model_channels) self.code_converter = nn.Sequential( - AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True), - AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True), - AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True), + AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True, tortoise_norm=True), + AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True, tortoise_norm=True), + AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True, tortoise_norm=True), ) self.code_norm = normalization(model_channels) self.latent_conditioner = nn.Sequential( nn.Conv1d(in_latent_channels, model_channels, 3, padding=1), - AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True), - AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True), - AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True), - AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True), + AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True, tortoise_norm=True), + AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True, tortoise_norm=True), + AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True, tortoise_norm=True), + AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True, tortoise_norm=True), ) self.contextual_embedder = nn.Sequential( nn.Conv1d(in_channels, model_channels, 3, padding=1, stride=2), @@ -197,31 +196,31 @@ def __init__( model_channels * 2, num_heads, relative_pos_embeddings=True, - do_checkpoint=False, + tortoise_norm=True, ), AttentionBlock( model_channels * 2, num_heads, relative_pos_embeddings=True, - do_checkpoint=False, + tortoise_norm=True, ), AttentionBlock( model_channels * 2, num_heads, relative_pos_embeddings=True, - do_checkpoint=False, + tortoise_norm=True, ), AttentionBlock( model_channels * 2, num_heads, relative_pos_embeddings=True, - do_checkpoint=False, + tortoise_norm=True, ), AttentionBlock( model_channels * 2, num_heads, relative_pos_embeddings=True, - do_checkpoint=False, + tortoise_norm=True, ), ) self.unconditioned_embedding = nn.Parameter(torch.randn(1, model_channels, 1)) @@ -385,7 +384,7 @@ def forward( unused_params.extend(list(lyr.parameters())) else: # First and last blocks will have autocast disabled for improved precision. - with autocast(x.device.type, enabled=self.enable_fp16 and i != 0): + with torch.autocast(x.device.type, enabled=self.enable_fp16 and i != 0): x = lyr(x, time_emb) x = x.float() diff --git a/TTS/tts/layers/tortoise/dpm_solver.py b/TTS/tts/layers/tortoise/dpm_solver.py index c70888df42..6a1d8ff784 100644 --- a/TTS/tts/layers/tortoise/dpm_solver.py +++ b/TTS/tts/layers/tortoise/dpm_solver.py @@ -1,7 +1,10 @@ +import logging import math import torch +logger = logging.getLogger(__name__) + class NoiseScheduleVP: def __init__( @@ -1171,7 +1174,7 @@ def norm_fn(v): lambda_0 - lambda_s, ) nfe += order - print("adaptive solver nfe", nfe) + logger.debug("adaptive solver nfe %d", nfe) return x def add_noise(self, x, t, noise=None): diff --git a/TTS/tts/layers/tortoise/transformer.py b/TTS/tts/layers/tortoise/transformer.py index 70d46aa3e0..ed4d79d4ab 100644 --- a/TTS/tts/layers/tortoise/transformer.py +++ b/TTS/tts/layers/tortoise/transformer.py @@ -1,22 +1,19 @@ +from typing import TypeVar, Union + import torch import torch.nn.functional as F from einops import rearrange from torch import nn -# helpers - +from TTS.utils.generic_utils import exists -def exists(val): - return val is not None - - -def default(val, d): - return val if exists(val) else d +# helpers +_T = TypeVar("_T") -def cast_tuple(val, depth=1): +def cast_tuple(val: Union[tuple[_T], list[_T], _T], depth: int = 1) -> tuple[_T]: if isinstance(val, list): - val = tuple(val) + return tuple(val) return val if isinstance(val, tuple) else (val,) * depth @@ -37,7 +34,7 @@ def route_args(router, args, depth): for key in matched_keys: val = args[key] for depth, ((f_args, g_args), routes) in enumerate(zip(routed_args, router[key])): - new_f_args, new_g_args = map(lambda route: ({key: val} if route else {}), routes) + new_f_args, new_g_args = (({key: val} if route else {}) for route in routes) routed_args[depth] = ({**f_args, **new_f_args}, {**g_args, **new_g_args}) return routed_args @@ -152,7 +149,7 @@ def forward(self, x, mask=None): softmax = torch.softmax qkv = self.to_qkv(x).chunk(3, dim=-1) - q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), qkv) + q, k, v = (rearrange(t, "b n (h d) -> b h n d", h=h) for t in qkv) q = q * self.scale diff --git a/TTS/tts/layers/tortoise/utils.py b/TTS/tts/layers/tortoise/utils.py index 810a9e7f7a..898121f793 100644 --- a/TTS/tts/layers/tortoise/utils.py +++ b/TTS/tts/layers/tortoise/utils.py @@ -1,8 +1,11 @@ +import logging import os from urllib import request from tqdm import tqdm +logger = logging.getLogger(__name__) + DEFAULT_MODELS_DIR = os.path.join(os.path.expanduser("~"), ".cache", "tortoise", "models") MODELS_DIR = os.environ.get("TORTOISE_MODELS_DIR", DEFAULT_MODELS_DIR) MODELS_DIR = "/data/speech_synth/models/" @@ -28,10 +31,10 @@ def download_models(specific_models=None): model_path = os.path.join(MODELS_DIR, model_name) if os.path.exists(model_path): continue - print(f"Downloading {model_name} from {url}...") + logger.info("Downloading %s from %s...", model_name, url) with tqdm(unit="B", unit_scale=True, unit_divisor=1024, miniters=1) as t: request.urlretrieve(url, model_path, lambda nb, bs, fs, t=t: t.update(nb * bs - t.n)) - print("Done.") + logger.info("Done.") def get_model_path(model_name, models_dir=MODELS_DIR): diff --git a/TTS/tts/layers/tortoise/xtransformers.py b/TTS/tts/layers/tortoise/xtransformers.py index 1eb3f77269..0892fee19d 100644 --- a/TTS/tts/layers/tortoise/xtransformers.py +++ b/TTS/tts/layers/tortoise/xtransformers.py @@ -1,13 +1,15 @@ import math from collections import namedtuple from functools import partial -from inspect import isfunction import torch import torch.nn.functional as F from einops import rearrange, repeat from torch import einsum, nn +from TTS.tts.layers.tortoise.transformer import cast_tuple, max_neg_value +from TTS.utils.generic_utils import default, exists + DEFAULT_DIM_HEAD = 64 Intermediates = namedtuple("Intermediates", ["pre_softmax_attn", "post_softmax_attn"]) @@ -25,20 +27,6 @@ # helpers -def exists(val): - return val is not None - - -def default(val, d): - if exists(val): - return val - return d() if isfunction(d) else d - - -def cast_tuple(val, depth): - return val if isinstance(val, tuple) else (val,) * depth - - class always: def __init__(self, val): self.val = val @@ -63,10 +51,6 @@ def __call__(self, x, *args, **kwargs): return x == self.val -def max_neg_value(tensor): - return -torch.finfo(tensor.dtype).max - - def l2norm(t): return F.normalize(t, p=2, dim=-1) @@ -84,7 +68,7 @@ def init_zero_(layer): def pick_and_pop(keys, d): - values = list(map(lambda key: d.pop(key), keys)) + values = [d.pop(key) for key in keys] return dict(zip(keys, values)) @@ -107,7 +91,7 @@ def group_by_key_prefix(prefix, d): def groupby_prefix_and_trim(prefix, d): kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d) - kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix) :], x[1]), tuple(kwargs_with_prefix.items()))) + kwargs_without_prefix = {x[0][len(prefix) :]: x[1] for x in tuple(kwargs_with_prefix.items())} return kwargs_without_prefix, kwargs @@ -428,7 +412,7 @@ def forward(self, x, **kwargs): feats_per_shift = x.shape[-1] // segments splitted = x.split(feats_per_shift, dim=-1) segments_to_shift, rest = splitted[:segments], splitted[segments:] - segments_to_shift = list(map(lambda args: shift(*args, mask=mask), zip(segments_to_shift, shifts))) + segments_to_shift = [shift(*args, mask=mask) for args in zip(segments_to_shift, shifts)] x = torch.cat((*segments_to_shift, *rest), dim=-1) return self.fn(x, **kwargs) @@ -635,7 +619,7 @@ def forward( v = self.to_v(v_input) if not collab_heads: - q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v)) + q, k, v = (rearrange(t, "b n (h d) -> b h n d", h=h) for t in (q, k, v)) else: q = einsum("b i d, h d -> b h i d", q, self.collab_mixing) k = rearrange(k, "b n d -> b () n d") @@ -650,9 +634,9 @@ def forward( if exists(rotary_pos_emb) and not has_context: l = rotary_pos_emb.shape[-1] - (ql, qr), (kl, kr), (vl, vr) = map(lambda t: (t[..., :l], t[..., l:]), (q, k, v)) - ql, kl, vl = map(lambda t: apply_rotary_pos_emb(t, rotary_pos_emb), (ql, kl, vl)) - q, k, v = map(lambda t: torch.cat(t, dim=-1), ((ql, qr), (kl, kr), (vl, vr))) + (ql, qr), (kl, kr), (vl, vr) = ((t[..., :l], t[..., l:]) for t in (q, k, v)) + ql, kl, vl = (apply_rotary_pos_emb(t, rotary_pos_emb) for t in (ql, kl, vl)) + q, k, v = (torch.cat(t, dim=-1) for t in ((ql, qr), (kl, kr), (vl, vr))) input_mask = None if any(map(exists, (mask, context_mask))): @@ -664,7 +648,7 @@ def forward( input_mask = q_mask * k_mask if self.num_mem_kv > 0: - mem_k, mem_v = map(lambda t: repeat(t, "h n d -> b h n d", b=b), (self.mem_k, self.mem_v)) + mem_k, mem_v = (repeat(t, "h n d -> b h n d", b=b) for t in (self.mem_k, self.mem_v)) k = torch.cat((mem_k, k), dim=-2) v = torch.cat((mem_v, v), dim=-2) if exists(input_mask): @@ -964,9 +948,7 @@ def forward( seq_len = x.shape[1] if past_key_values is not None: seq_len += past_key_values[0][0].shape[-2] - max_rotary_emb_length = max( - list(map(lambda m: (m.shape[1] if exists(m) else 0) + seq_len, mems)) + [expected_seq_len] - ) + max_rotary_emb_length = max([(m.shape[1] if exists(m) else 0) + seq_len for m in mems] + [expected_seq_len]) rotary_pos_emb = self.rotary_pos_emb(max_rotary_emb_length, x.device) present_key_values = [] @@ -1200,7 +1182,7 @@ def forward( res = [out] if return_attn: - attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates)) + attn_maps = [t.post_softmax_attn for t in intermediates.attn_intermediates] res.append(attn_maps) if use_cache: res.append(intermediates.past_key_values) @@ -1249,7 +1231,7 @@ def forward(self, x, return_embeddings=False, mask=None, return_attn=False, mems res = [out] if return_attn: - attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates)) + attn_maps = [t.post_softmax_attn for t in intermediates.attn_intermediates] res.append(attn_maps) if use_cache: res.append(intermediates.past_key_values) diff --git a/TTS/tts/layers/vits/discriminator.py b/TTS/tts/layers/vits/discriminator.py index c27d11bef6..49f7a0d074 100644 --- a/TTS/tts/layers/vits/discriminator.py +++ b/TTS/tts/layers/vits/discriminator.py @@ -2,7 +2,7 @@ from torch import nn from torch.nn.modules.conv import Conv1d -from TTS.vocoder.models.hifigan_discriminator import DiscriminatorP, MultiPeriodDiscriminator +from TTS.vocoder.models.hifigan_discriminator import LRELU_SLOPE, DiscriminatorP class DiscriminatorS(torch.nn.Module): @@ -39,7 +39,7 @@ def forward(self, x): feat = [] for l in self.convs: x = l(x) - x = torch.nn.functional.leaky_relu(x, 0.1) + x = torch.nn.functional.leaky_relu(x, LRELU_SLOPE) feat.append(x) x = self.conv_post(x) feat.append(x) diff --git a/TTS/tts/layers/vits/networks.py b/TTS/tts/layers/vits/networks.py index f97b584fe6..ab2ca5667a 100644 --- a/TTS/tts/layers/vits/networks.py +++ b/TTS/tts/layers/vits/networks.py @@ -10,22 +10,6 @@ LRELU_SLOPE = 0.1 -def convert_pad_shape(pad_shape): - l = pad_shape[::-1] - pad_shape = [item for sublist in l for item in sublist] - return pad_shape - - -def init_weights(m, mean=0.0, std=0.01): - classname = m.__class__.__name__ - if classname.find("Conv") != -1: - m.weight.data.normal_(mean, std) - - -def get_padding(kernel_size, dilation=1): - return int((kernel_size * dilation - dilation) / 2) - - class TextEncoder(nn.Module): def __init__( self, @@ -272,7 +256,7 @@ def __init__( ) self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) - def forward(self, x, x_lengths, g=None): + def forward(self, x, x_lengths, g=None, tau=1.0): """ Shapes: - x: :math:`[B, C, T]` @@ -284,5 +268,5 @@ def forward(self, x, x_lengths, g=None): x = self.enc(x, x_mask, g=g) stats = self.proj(x) * x_mask mean, log_scale = torch.split(stats, self.out_channels, dim=1) - z = (mean + torch.randn_like(mean) * torch.exp(log_scale)) * x_mask + z = (mean + torch.randn_like(mean) * tau * torch.exp(log_scale)) * x_mask return z, mean, log_scale, x_mask diff --git a/TTS/tts/utils/monotonic_align/__init__.py b/TTS/tts/layers/xtts/__init__.py similarity index 100% rename from TTS/tts/utils/monotonic_align/__init__.py rename to TTS/tts/layers/xtts/__init__.py diff --git a/TTS/tts/layers/xtts/dvae.py b/TTS/tts/layers/xtts/dvae.py index bdd7a9d09f..4f806f82cb 100644 --- a/TTS/tts/layers/xtts/dvae.py +++ b/TTS/tts/layers/xtts/dvae.py @@ -1,4 +1,5 @@ import functools +import logging from math import sqrt import torch @@ -8,9 +9,9 @@ import torchaudio from einops import rearrange +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 -def default(val, d): - return val if val is not None else d +logger = logging.getLogger(__name__) def eval_decorator(fn): @@ -43,7 +44,7 @@ def dvae_wav_to_mel( mel = mel_stft(wav) mel = torch.log(torch.clamp(mel, min=1e-5)) if mel_norms is None: - mel_norms = torch.load(mel_norms_file, map_location=device) + mel_norms = torch.load(mel_norms_file, map_location=device, weights_only=is_pytorch_at_least_2_4()) mel = mel / mel_norms.unsqueeze(0).unsqueeze(-1) return mel @@ -79,7 +80,7 @@ def forward(self, input, return_soft_codes=False): self.embed_avg = (ea * ~mask + rand_embed).permute(1, 0) self.cluster_size = self.cluster_size * ~mask.squeeze() if torch.any(mask): - print(f"Reset {torch.sum(mask)} embedding codes.") + logger.info("Reset %d embedding codes.", torch.sum(mask)) self.codes = None self.codes_full = False @@ -260,7 +261,7 @@ def __init__( dec_init_chan = codebook_dim if not has_resblocks else dec_chans[0] dec_chans = [dec_init_chan, *dec_chans] - enc_chans_io, dec_chans_io = map(lambda t: list(zip(t[:-1], t[1:])), (enc_chans, dec_chans)) + enc_chans_io, dec_chans_io = (list(zip(t[:-1], t[1:])) for t in (enc_chans, dec_chans)) pad = (kernel_size - 1) // 2 for (enc_in, enc_out), (dec_in, dec_out) in zip(enc_chans_io, dec_chans_io): @@ -306,9 +307,9 @@ def norm(self, images): if not self.normalization is not None: return images - means, stds = map(lambda t: torch.as_tensor(t).to(images), self.normalization) + means, stds = (torch.as_tensor(t).to(images) for t in self.normalization) arrange = "c -> () c () ()" if self.positional_dims == 2 else "c -> () c ()" - means, stds = map(lambda t: rearrange(t, arrange), (means, stds)) + means, stds = (rearrange(t, arrange) for t in (means, stds)) images = images.clone() images.sub_(means).div_(stds) return images diff --git a/TTS/tts/layers/xtts/gpt.py b/TTS/tts/layers/xtts/gpt.py index e7b186b858..20eff26ecc 100644 --- a/TTS/tts/layers/xtts/gpt.py +++ b/TTS/tts/layers/xtts/gpt.py @@ -1,7 +1,5 @@ # ported from: https://github.com/neonbjb/tortoise-tts -import functools -import math import random import torch @@ -9,82 +7,16 @@ import torch.nn.functional as F from transformers import GPT2Config +from TTS.tts.layers.tortoise.autoregressive import ( + ConditioningEncoder, + LearnedPositionEmbeddings, + _prepare_attention_mask_for_generation, + build_hf_gpt_transformer, +) from TTS.tts.layers.xtts.gpt_inference import GPT2InferenceModel -from TTS.tts.layers.xtts.latent_encoder import ConditioningEncoder from TTS.tts.layers.xtts.perceiver_encoder import PerceiverResampler -def null_position_embeddings(range, dim): - return torch.zeros((range.shape[0], range.shape[1], dim), device=range.device) - - -class LearnedPositionEmbeddings(nn.Module): - def __init__(self, seq_len, model_dim, init=0.02, relative=False): - super().__init__() - # nn.Embedding - self.emb = torch.nn.Embedding(seq_len, model_dim) - # Initializing this way is standard for GPT-2 - self.emb.weight.data.normal_(mean=0.0, std=init) - self.relative = relative - self.seq_len = seq_len - - def forward(self, x): - sl = x.shape[1] - if self.relative: - start = random.randint(sl, self.seq_len) - sl - return self.emb(torch.arange(start, start + sl, device=x.device)) - else: - return self.emb(torch.arange(0, sl, device=x.device)) - - def get_fixed_embedding(self, ind, dev): - return self.emb(torch.tensor([ind], device=dev)).unsqueeze(0) - - -def build_hf_gpt_transformer( - layers, - model_dim, - heads, - max_mel_seq_len, - max_text_seq_len, - max_prompt_len, - checkpointing, -): - """ - GPT-2 implemented by the HuggingFace library. - """ - from transformers import GPT2Config, GPT2Model - - gpt_config = GPT2Config( - vocab_size=256, # Unused. - n_positions=max_mel_seq_len + max_text_seq_len + max_prompt_len, - n_ctx=max_mel_seq_len + max_text_seq_len + max_prompt_len, - n_embd=model_dim, - n_layer=layers, - n_head=heads, - gradient_checkpointing=checkpointing, - use_cache=not checkpointing, - ) - gpt = GPT2Model(gpt_config) - # Override the built in positional embeddings - del gpt.wpe - gpt.wpe = functools.partial(null_position_embeddings, dim=model_dim) - # Built-in token embeddings are unused. - del gpt.wte - - mel_pos_emb = ( - LearnedPositionEmbeddings(max_mel_seq_len, model_dim) - if max_mel_seq_len != -1 - else functools.partial(null_position_embeddings, dim=model_dim) - ) - text_pos_emb = ( - LearnedPositionEmbeddings(max_text_seq_len, model_dim) - if max_mel_seq_len != -1 - else functools.partial(null_position_embeddings, dim=model_dim) - ) - # gpt = torch.compile(gpt, mode="reduce-overhead", fullgraph=True) - return gpt, mel_pos_emb, text_pos_emb, None, None - - class GPT(nn.Module): def __init__( self, @@ -149,13 +81,13 @@ def __init__( self.mel_layer_pos_embedding, self.text_layer_pos_embedding, ) = build_hf_gpt_transformer( - layers, - model_dim, - heads, - self.max_mel_tokens, - self.max_text_tokens, - self.max_prompt_tokens, - checkpointing, + layers=layers, + model_dim=model_dim, + heads=heads, + max_mel_seq_len=self.max_mel_tokens, + max_text_seq_len=self.max_text_tokens, + max_prompt_len=self.max_prompt_tokens, + checkpointing=checkpointing, ) if train_solo_embeddings: self.mel_solo_embedding = nn.Parameter(torch.randn(1, 1, model_dim) * 0.02, requires_grad=True) @@ -188,9 +120,9 @@ def __init__( def get_grad_norm_parameter_groups(self): return { "conditioning_encoder": list(self.conditioning_encoder.parameters()), - "conditioning_perceiver": list(self.conditioning_perceiver.parameters()) - if self.use_perceiver_resampler - else None, + "conditioning_perceiver": ( + list(self.conditioning_perceiver.parameters()) if self.use_perceiver_resampler else None + ), "gpt": list(self.gpt.parameters()), "heads": list(self.text_head.parameters()) + list(self.mel_head.parameters()), } @@ -303,19 +235,6 @@ def get_logits( else: return first_logits - def get_conditioning(self, speech_conditioning_input): - speech_conditioning_input = ( - speech_conditioning_input.unsqueeze(1) - if len(speech_conditioning_input.shape) == 3 - else speech_conditioning_input - ) - conds = [] - for j in range(speech_conditioning_input.shape[1]): - conds.append(self.conditioning_encoder(speech_conditioning_input[:, j])) - conds = torch.stack(conds, dim=1) - conds = conds.mean(dim=1) - return conds - def get_prompts(self, prompt_codes): """ Create a prompt from the mel codes. This is used to condition the model on the mel codes. @@ -354,6 +273,7 @@ def get_style_emb(self, cond_input, return_latent=False): """ cond_input: (b, 80, s) or (b, 1, 80, s) conds: (b, 1024, s) + output: (b, 1024, 32) """ conds = None if not return_latent: @@ -587,12 +507,15 @@ def generate( **hf_generate_kwargs, ): gpt_inputs = self.compute_embeddings(cond_latents, text_inputs) + stop_token_tensor = torch.tensor(self.stop_audio_token, device=gpt_inputs.device, dtype=torch.long) + attention_mask = _prepare_attention_mask_for_generation(gpt_inputs, stop_token_tensor, stop_token_tensor) gen = self.gpt_inference.generate( gpt_inputs, bos_token_id=self.start_audio_token, pad_token_id=self.stop_audio_token, eos_token_id=self.stop_audio_token, max_length=self.max_gen_mel_tokens + gpt_inputs.shape[-1], + attention_mask=attention_mask, **hf_generate_kwargs, ) if "return_dict_in_generate" in hf_generate_kwargs: diff --git a/TTS/tts/layers/xtts/gpt_inference.py b/TTS/tts/layers/xtts/gpt_inference.py index d44bd3decd..e94683524a 100644 --- a/TTS/tts/layers/xtts/gpt_inference.py +++ b/TTS/tts/layers/xtts/gpt_inference.py @@ -1,12 +1,12 @@ -import math - import torch from torch import nn -from transformers import GPT2PreTrainedModel +from transformers import GenerationMixin, GPT2PreTrainedModel from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions +from TTS.tts.layers.xtts.stream_generator import StreamGenerationConfig + -class GPT2InferenceModel(GPT2PreTrainedModel): +class GPT2InferenceModel(GPT2PreTrainedModel, GenerationMixin): """Override GPT2LMHeadModel to allow for prefix conditioning.""" def __init__(self, config, gpt, pos_emb, embeddings, norm, linear, kv_cache): @@ -17,6 +17,7 @@ def __init__(self, config, gpt, pos_emb, embeddings, norm, linear, kv_cache): self.final_norm = norm self.lm_head = nn.Sequential(norm, linear) self.kv_cache = kv_cache + self.generation_config = StreamGenerationConfig.from_model_config(config) if self.can_generate() else None def store_prefix_emb(self, prefix_emb): self.cached_prefix_emb = prefix_emb diff --git a/TTS/tts/layers/xtts/hifigan_decoder.py b/TTS/tts/layers/xtts/hifigan_decoder.py index 9add7826e6..2e6ac01a87 100644 --- a/TTS/tts/layers/xtts/hifigan_decoder.py +++ b/TTS/tts/layers/xtts/hifigan_decoder.py @@ -1,615 +1,12 @@ -import torch -import torchaudio -from torch import nn -from torch.nn import Conv1d, ConvTranspose1d -from torch.nn import functional as F -from torch.nn.utils.parametrizations import weight_norm -from torch.nn.utils.parametrize import remove_parametrizations - -from TTS.utils.io import load_fsspec - -LRELU_SLOPE = 0.1 - - -def get_padding(k, d): - return int((k * d - d) / 2) - - -class ResBlock1(torch.nn.Module): - """Residual Block Type 1. It has 3 convolutional layers in each convolutional block. - - Network:: - - x -> lrelu -> conv1_1 -> conv1_2 -> conv1_3 -> z -> lrelu -> conv2_1 -> conv2_2 -> conv2_3 -> o -> + -> o - |--------------------------------------------------------------------------------------------------| - - - Args: - channels (int): number of hidden channels for the convolutional layers. - kernel_size (int): size of the convolution filter in each layer. - dilations (list): list of dilation value for each conv layer in a block. - """ - - def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)): - super().__init__() - self.convs1 = nn.ModuleList( - [ - weight_norm( - Conv1d( - channels, - channels, - kernel_size, - 1, - dilation=dilation[0], - padding=get_padding(kernel_size, dilation[0]), - ) - ), - weight_norm( - Conv1d( - channels, - channels, - kernel_size, - 1, - dilation=dilation[1], - padding=get_padding(kernel_size, dilation[1]), - ) - ), - weight_norm( - Conv1d( - channels, - channels, - kernel_size, - 1, - dilation=dilation[2], - padding=get_padding(kernel_size, dilation[2]), - ) - ), - ] - ) - - self.convs2 = nn.ModuleList( - [ - weight_norm( - Conv1d( - channels, - channels, - kernel_size, - 1, - dilation=1, - padding=get_padding(kernel_size, 1), - ) - ), - weight_norm( - Conv1d( - channels, - channels, - kernel_size, - 1, - dilation=1, - padding=get_padding(kernel_size, 1), - ) - ), - weight_norm( - Conv1d( - channels, - channels, - kernel_size, - 1, - dilation=1, - padding=get_padding(kernel_size, 1), - ) - ), - ] - ) - - def forward(self, x): - """ - Args: - x (Tensor): input tensor. - Returns: - Tensor: output tensor. - Shapes: - x: [B, C, T] - """ - for c1, c2 in zip(self.convs1, self.convs2): - xt = F.leaky_relu(x, LRELU_SLOPE) - xt = c1(xt) - xt = F.leaky_relu(xt, LRELU_SLOPE) - xt = c2(xt) - x = xt + x - return x - - def remove_weight_norm(self): - for l in self.convs1: - remove_parametrizations(l, "weight") - for l in self.convs2: - remove_parametrizations(l, "weight") - - -class ResBlock2(torch.nn.Module): - """Residual Block Type 2. It has 1 convolutional layers in each convolutional block. - - Network:: - - x -> lrelu -> conv1-> -> z -> lrelu -> conv2-> o -> + -> o - |---------------------------------------------------| - - - Args: - channels (int): number of hidden channels for the convolutional layers. - kernel_size (int): size of the convolution filter in each layer. - dilations (list): list of dilation value for each conv layer in a block. - """ - - def __init__(self, channels, kernel_size=3, dilation=(1, 3)): - super().__init__() - self.convs = nn.ModuleList( - [ - weight_norm( - Conv1d( - channels, - channels, - kernel_size, - 1, - dilation=dilation[0], - padding=get_padding(kernel_size, dilation[0]), - ) - ), - weight_norm( - Conv1d( - channels, - channels, - kernel_size, - 1, - dilation=dilation[1], - padding=get_padding(kernel_size, dilation[1]), - ) - ), - ] - ) - - def forward(self, x): - for c in self.convs: - xt = F.leaky_relu(x, LRELU_SLOPE) - xt = c(xt) - x = xt + x - return x - - def remove_weight_norm(self): - for l in self.convs: - remove_parametrizations(l, "weight") - - -class HifiganGenerator(torch.nn.Module): - def __init__( - self, - in_channels, - out_channels, - resblock_type, - resblock_dilation_sizes, - resblock_kernel_sizes, - upsample_kernel_sizes, - upsample_initial_channel, - upsample_factors, - inference_padding=5, - cond_channels=0, - conv_pre_weight_norm=True, - conv_post_weight_norm=True, - conv_post_bias=True, - cond_in_each_up_layer=False, - ): - r"""HiFiGAN Generator with Multi-Receptive Field Fusion (MRF) - - Network: - x -> lrelu -> upsampling_layer -> resblock1_k1x1 -> z1 -> + -> z_sum / #resblocks -> lrelu -> conv_post_7x1 -> tanh -> o - .. -> zI ---| - resblockN_kNx1 -> zN ---' - - Args: - in_channels (int): number of input tensor channels. - out_channels (int): number of output tensor channels. - resblock_type (str): type of the `ResBlock`. '1' or '2'. - resblock_dilation_sizes (List[List[int]]): list of dilation values in each layer of a `ResBlock`. - resblock_kernel_sizes (List[int]): list of kernel sizes for each `ResBlock`. - upsample_kernel_sizes (List[int]): list of kernel sizes for each transposed convolution. - upsample_initial_channel (int): number of channels for the first upsampling layer. This is divided by 2 - for each consecutive upsampling layer. - upsample_factors (List[int]): upsampling factors (stride) for each upsampling layer. - inference_padding (int): constant padding applied to the input at inference time. Defaults to 5. - """ - super().__init__() - self.inference_padding = inference_padding - self.num_kernels = len(resblock_kernel_sizes) - self.num_upsamples = len(upsample_factors) - self.cond_in_each_up_layer = cond_in_each_up_layer - - # initial upsampling layers - self.conv_pre = weight_norm(Conv1d(in_channels, upsample_initial_channel, 7, 1, padding=3)) - resblock = ResBlock1 if resblock_type == "1" else ResBlock2 - # upsampling layers - self.ups = nn.ModuleList() - for i, (u, k) in enumerate(zip(upsample_factors, upsample_kernel_sizes)): - self.ups.append( - weight_norm( - ConvTranspose1d( - upsample_initial_channel // (2**i), - upsample_initial_channel // (2 ** (i + 1)), - k, - u, - padding=(k - u) // 2, - ) - ) - ) - # MRF blocks - self.resblocks = nn.ModuleList() - for i in range(len(self.ups)): - ch = upsample_initial_channel // (2 ** (i + 1)) - for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)): - self.resblocks.append(resblock(ch, k, d)) - # post convolution layer - self.conv_post = weight_norm(Conv1d(ch, out_channels, 7, 1, padding=3, bias=conv_post_bias)) - if cond_channels > 0: - self.cond_layer = nn.Conv1d(cond_channels, upsample_initial_channel, 1) - - if not conv_pre_weight_norm: - remove_parametrizations(self.conv_pre, "weight") - - if not conv_post_weight_norm: - remove_parametrizations(self.conv_post, "weight") - - if self.cond_in_each_up_layer: - self.conds = nn.ModuleList() - for i in range(len(self.ups)): - ch = upsample_initial_channel // (2 ** (i + 1)) - self.conds.append(nn.Conv1d(cond_channels, ch, 1)) - - def forward(self, x, g=None): - """ - Args: - x (Tensor): feature input tensor. - g (Tensor): global conditioning input tensor. - - Returns: - Tensor: output waveform. - - Shapes: - x: [B, C, T] - Tensor: [B, 1, T] - """ - o = self.conv_pre(x) - if hasattr(self, "cond_layer"): - o = o + self.cond_layer(g) - for i in range(self.num_upsamples): - o = F.leaky_relu(o, LRELU_SLOPE) - o = self.ups[i](o) - - if self.cond_in_each_up_layer: - o = o + self.conds[i](g) - - z_sum = None - for j in range(self.num_kernels): - if z_sum is None: - z_sum = self.resblocks[i * self.num_kernels + j](o) - else: - z_sum += self.resblocks[i * self.num_kernels + j](o) - o = z_sum / self.num_kernels - o = F.leaky_relu(o) - o = self.conv_post(o) - o = torch.tanh(o) - return o - - @torch.no_grad() - def inference(self, c): - """ - Args: - x (Tensor): conditioning input tensor. - - Returns: - Tensor: output waveform. - - Shapes: - x: [B, C, T] - Tensor: [B, 1, T] - """ - c = c.to(self.conv_pre.weight.device) - c = torch.nn.functional.pad(c, (self.inference_padding, self.inference_padding), "replicate") - return self.forward(c) - - def remove_weight_norm(self): - print("Removing weight norm...") - for l in self.ups: - remove_parametrizations(l, "weight") - for l in self.resblocks: - l.remove_weight_norm() - remove_parametrizations(self.conv_pre, "weight") - remove_parametrizations(self.conv_post, "weight") - - def load_checkpoint( - self, config, checkpoint_path, eval=False, cache=False - ): # pylint: disable=unused-argument, redefined-builtin - state = torch.load(checkpoint_path, map_location=torch.device("cpu")) - self.load_state_dict(state["model"]) - if eval: - self.eval() - assert not self.training - self.remove_weight_norm() - - -class SELayer(nn.Module): - def __init__(self, channel, reduction=8): - super(SELayer, self).__init__() - self.avg_pool = nn.AdaptiveAvgPool2d(1) - self.fc = nn.Sequential( - nn.Linear(channel, channel // reduction), - nn.ReLU(inplace=True), - nn.Linear(channel // reduction, channel), - nn.Sigmoid(), - ) - - def forward(self, x): - b, c, _, _ = x.size() - y = self.avg_pool(x).view(b, c) - y = self.fc(y).view(b, c, 1, 1) - return x * y - - -class SEBasicBlock(nn.Module): - expansion = 1 - - def __init__(self, inplanes, planes, stride=1, downsample=None, reduction=8): - super(SEBasicBlock, self).__init__() - self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=3, stride=stride, padding=1, bias=False) - self.bn1 = nn.BatchNorm2d(planes) - self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1, bias=False) - self.bn2 = nn.BatchNorm2d(planes) - self.relu = nn.ReLU(inplace=True) - self.se = SELayer(planes, reduction) - self.downsample = downsample - self.stride = stride - - def forward(self, x): - residual = x - - out = self.conv1(x) - out = self.relu(out) - out = self.bn1(out) - - out = self.conv2(out) - out = self.bn2(out) - out = self.se(out) - - if self.downsample is not None: - residual = self.downsample(x) - - out += residual - out = self.relu(out) - return out - - -def set_init_dict(model_dict, checkpoint_state, c): - # Partial initialization: if there is a mismatch with new and old layer, it is skipped. - for k, v in checkpoint_state.items(): - if k not in model_dict: - print(" | > Layer missing in the model definition: {}".format(k)) - # 1. filter out unnecessary keys - pretrained_dict = {k: v for k, v in checkpoint_state.items() if k in model_dict} - # 2. filter out different size layers - pretrained_dict = {k: v for k, v in pretrained_dict.items() if v.numel() == model_dict[k].numel()} - # 3. skip reinit layers - if c.has("reinit_layers") and c.reinit_layers is not None: - for reinit_layer_name in c.reinit_layers: - pretrained_dict = {k: v for k, v in pretrained_dict.items() if reinit_layer_name not in k} - # 4. overwrite entries in the existing state dict - model_dict.update(pretrained_dict) - print(" | > {} / {} layers are restored.".format(len(pretrained_dict), len(model_dict))) - return model_dict - - -class PreEmphasis(nn.Module): - def __init__(self, coefficient=0.97): - super().__init__() - self.coefficient = coefficient - self.register_buffer("filter", torch.FloatTensor([-self.coefficient, 1.0]).unsqueeze(0).unsqueeze(0)) - - def forward(self, x): - assert len(x.size()) == 2 - - x = torch.nn.functional.pad(x.unsqueeze(1), (1, 0), "reflect") - return torch.nn.functional.conv1d(x, self.filter).squeeze(1) - - -class ResNetSpeakerEncoder(nn.Module): - """This is copied from 🐸TTS to remove it from the dependencies.""" - - # pylint: disable=W0102 - def __init__( - self, - input_dim=64, - proj_dim=512, - layers=[3, 4, 6, 3], - num_filters=[32, 64, 128, 256], - encoder_type="ASP", - log_input=False, - use_torch_spec=False, - audio_config=None, - ): - super(ResNetSpeakerEncoder, self).__init__() - - self.encoder_type = encoder_type - self.input_dim = input_dim - self.log_input = log_input - self.use_torch_spec = use_torch_spec - self.audio_config = audio_config - self.proj_dim = proj_dim - - self.conv1 = nn.Conv2d(1, num_filters[0], kernel_size=3, stride=1, padding=1) - self.relu = nn.ReLU(inplace=True) - self.bn1 = nn.BatchNorm2d(num_filters[0]) - - self.inplanes = num_filters[0] - self.layer1 = self.create_layer(SEBasicBlock, num_filters[0], layers[0]) - self.layer2 = self.create_layer(SEBasicBlock, num_filters[1], layers[1], stride=(2, 2)) - self.layer3 = self.create_layer(SEBasicBlock, num_filters[2], layers[2], stride=(2, 2)) - self.layer4 = self.create_layer(SEBasicBlock, num_filters[3], layers[3], stride=(2, 2)) - - self.instancenorm = nn.InstanceNorm1d(input_dim) +import logging - if self.use_torch_spec: - self.torch_spec = torch.nn.Sequential( - PreEmphasis(audio_config["preemphasis"]), - torchaudio.transforms.MelSpectrogram( - sample_rate=audio_config["sample_rate"], - n_fft=audio_config["fft_size"], - win_length=audio_config["win_length"], - hop_length=audio_config["hop_length"], - window_fn=torch.hamming_window, - n_mels=audio_config["num_mels"], - ), - ) - - else: - self.torch_spec = None - - outmap_size = int(self.input_dim / 8) - - self.attention = nn.Sequential( - nn.Conv1d(num_filters[3] * outmap_size, 128, kernel_size=1), - nn.ReLU(), - nn.BatchNorm1d(128), - nn.Conv1d(128, num_filters[3] * outmap_size, kernel_size=1), - nn.Softmax(dim=2), - ) - - if self.encoder_type == "SAP": - out_dim = num_filters[3] * outmap_size - elif self.encoder_type == "ASP": - out_dim = num_filters[3] * outmap_size * 2 - else: - raise ValueError("Undefined encoder") - - self.fc = nn.Linear(out_dim, proj_dim) - - self._init_layers() - - def _init_layers(self): - for m in self.modules(): - if isinstance(m, nn.Conv2d): - nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu") - elif isinstance(m, nn.BatchNorm2d): - nn.init.constant_(m.weight, 1) - nn.init.constant_(m.bias, 0) - - def create_layer(self, block, planes, blocks, stride=1): - downsample = None - if stride != 1 or self.inplanes != planes * block.expansion: - downsample = nn.Sequential( - nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), - nn.BatchNorm2d(planes * block.expansion), - ) - - layers = [] - layers.append(block(self.inplanes, planes, stride, downsample)) - self.inplanes = planes * block.expansion - for _ in range(1, blocks): - layers.append(block(self.inplanes, planes)) - - return nn.Sequential(*layers) - - # pylint: disable=R0201 - def new_parameter(self, *size): - out = nn.Parameter(torch.FloatTensor(*size)) - nn.init.xavier_normal_(out) - return out - - def forward(self, x, l2_norm=False): - """Forward pass of the model. - - Args: - x (Tensor): Raw waveform signal or spectrogram frames. If input is a waveform, `torch_spec` must be `True` - to compute the spectrogram on-the-fly. - l2_norm (bool): Whether to L2-normalize the outputs. - - Shapes: - - x: :math:`(N, 1, T_{in})` or :math:`(N, D_{spec}, T_{in})` - """ - x.squeeze_(1) - # if you torch spec compute it otherwise use the mel spec computed by the AP - if self.use_torch_spec: - x = self.torch_spec(x) - - if self.log_input: - x = (x + 1e-6).log() - x = self.instancenorm(x).unsqueeze(1) - - x = self.conv1(x) - x = self.relu(x) - x = self.bn1(x) - - x = self.layer1(x) - x = self.layer2(x) - x = self.layer3(x) - x = self.layer4(x) - - x = x.reshape(x.size()[0], -1, x.size()[-1]) - - w = self.attention(x) - - if self.encoder_type == "SAP": - x = torch.sum(x * w, dim=2) - elif self.encoder_type == "ASP": - mu = torch.sum(x * w, dim=2) - sg = torch.sqrt((torch.sum((x**2) * w, dim=2) - mu**2).clamp(min=1e-5)) - x = torch.cat((mu, sg), 1) - - x = x.view(x.size()[0], -1) - x = self.fc(x) - - if l2_norm: - x = torch.nn.functional.normalize(x, p=2, dim=1) - return x - - def load_checkpoint( - self, - checkpoint_path: str, - eval: bool = False, - use_cuda: bool = False, - criterion=None, - cache=False, - ): - state = load_fsspec(checkpoint_path, map_location=torch.device("cpu"), cache=cache) - try: - self.load_state_dict(state["model"]) - print(" > Model fully restored. ") - except (KeyError, RuntimeError) as error: - # If eval raise the error - if eval: - raise error - - print(" > Partial model initialization.") - model_dict = self.state_dict() - model_dict = set_init_dict(model_dict, state["model"]) - self.load_state_dict(model_dict) - del model_dict - - # load the criterion for restore_path - if criterion is not None and "criterion" in state: - try: - criterion.load_state_dict(state["criterion"]) - except (KeyError, RuntimeError) as error: - print(" > Criterion load ignored because of:", error) - - if use_cuda: - self.cuda() - if criterion is not None: - criterion = criterion.cuda() +import torch +from trainer.io import load_fsspec - if eval: - self.eval() - assert not self.training +from TTS.encoder.models.resnet import ResNetSpeakerEncoder +from TTS.vocoder.models.hifigan_generator import HifiganGenerator - if not eval: - return criterion, state["step"] - return criterion +logger = logging.getLogger(__name__) class HifiDecoder(torch.nn.Module): diff --git a/TTS/tts/layers/xtts/latent_encoder.py b/TTS/tts/layers/xtts/latent_encoder.py deleted file mode 100644 index f9d62a36f1..0000000000 --- a/TTS/tts/layers/xtts/latent_encoder.py +++ /dev/null @@ -1,141 +0,0 @@ -# ported from: Originally ported from: https://github.com/neonbjb/tortoise-tts - -import math - -import torch -from torch import nn -from torch.nn import functional as F - - -class GroupNorm32(nn.GroupNorm): - def forward(self, x): - return super().forward(x.float()).type(x.dtype) - - -def conv_nd(dims, *args, **kwargs): - if dims == 1: - return nn.Conv1d(*args, **kwargs) - elif dims == 2: - return nn.Conv2d(*args, **kwargs) - elif dims == 3: - return nn.Conv3d(*args, **kwargs) - raise ValueError(f"unsupported dimensions: {dims}") - - -def normalization(channels): - groups = 32 - if channels <= 16: - groups = 8 - elif channels <= 64: - groups = 16 - while channels % groups != 0: - groups = int(groups / 2) - assert groups > 2 - return GroupNorm32(groups, channels) - - -def zero_module(module): - for p in module.parameters(): - p.detach().zero_() - return module - - -class QKVAttention(nn.Module): - def __init__(self, n_heads): - super().__init__() - self.n_heads = n_heads - - def forward(self, qkv, mask=None, qk_bias=0): - """ - Apply QKV attention. - - :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. - :return: an [N x (H * C) x T] tensor after attention. - """ - bs, width, length = qkv.shape - assert width % (3 * self.n_heads) == 0 - ch = width // (3 * self.n_heads) - q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) - scale = 1 / math.sqrt(math.sqrt(ch)) - weight = torch.einsum("bct,bcs->bts", q * scale, k * scale) # More stable with f16 than dividing afterwards - weight = weight + qk_bias - if mask is not None: - mask = mask.repeat(self.n_heads, 1, 1) - weight[mask.logical_not()] = -torch.inf - weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) - a = torch.einsum("bts,bcs->bct", weight, v) - - return a.reshape(bs, -1, length) - - -class AttentionBlock(nn.Module): - """An attention block that allows spatial positions to attend to each other.""" - - def __init__( - self, - channels, - num_heads=1, - num_head_channels=-1, - out_channels=None, - do_activation=False, - ): - super().__init__() - self.channels = channels - out_channels = channels if out_channels is None else out_channels - self.do_activation = do_activation - if num_head_channels == -1: - self.num_heads = num_heads - else: - assert ( - channels % num_head_channels == 0 - ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" - self.num_heads = channels // num_head_channels - self.norm = normalization(channels) - self.qkv = conv_nd(1, channels, out_channels * 3, 1) - self.attention = QKVAttention(self.num_heads) - - self.x_proj = nn.Identity() if out_channels == channels else conv_nd(1, channels, out_channels, 1) - self.proj_out = zero_module(conv_nd(1, out_channels, out_channels, 1)) - - def forward(self, x, mask=None, qk_bias=0): - b, c, *spatial = x.shape - if mask is not None: - if len(mask.shape) == 2: - mask = mask.unsqueeze(0).repeat(x.shape[0], 1, 1) - if mask.shape[1] != x.shape[-1]: - mask = mask[:, : x.shape[-1], : x.shape[-1]] - - x = x.reshape(b, c, -1) - x = self.norm(x) - if self.do_activation: - x = F.silu(x, inplace=True) - qkv = self.qkv(x) - h = self.attention(qkv, mask=mask, qk_bias=qk_bias) - h = self.proj_out(h) - xp = self.x_proj(x) - return (xp + h).reshape(b, xp.shape[1], *spatial) - - -class ConditioningEncoder(nn.Module): - def __init__( - self, - spec_dim, - embedding_dim, - attn_blocks=6, - num_attn_heads=4, - ): - super().__init__() - attn = [] - self.init = nn.Conv1d(spec_dim, embedding_dim, kernel_size=1) - for a in range(attn_blocks): - attn.append(AttentionBlock(embedding_dim, num_attn_heads)) - self.attn = nn.Sequential(*attn) - self.dim = embedding_dim - - def forward(self, x): - """ - x: (b, 80, s) - """ - h = self.init(x) - h = self.attn(h) - return h diff --git a/TTS/tts/layers/xtts/perceiver_encoder.py b/TTS/tts/layers/xtts/perceiver_encoder.py index 7b7ee79b50..7477087283 100644 --- a/TTS/tts/layers/xtts/perceiver_encoder.py +++ b/TTS/tts/layers/xtts/perceiver_encoder.py @@ -7,12 +7,10 @@ import torch.nn.functional as F from einops import rearrange, repeat from einops.layers.torch import Rearrange -from packaging import version from torch import einsum, nn - -def exists(val): - return val is not None +from TTS.tts.layers.tortoise.transformer import GEGLU +from TTS.utils.generic_utils import default, exists def once(fn): @@ -44,9 +42,6 @@ def __init__(self, dropout=0.0, causal=False, use_flash=False): self.register_buffer("mask", None, persistent=False) self.use_flash = use_flash - assert not ( - use_flash and version.parse(torch.__version__) < version.parse("2.0.0") - ), "in order to use flash attention, you must be using pytorch 2.0 or above" # determine efficient attention configs for cuda and cpu self.config = namedtuple("EfficientAttentionConfig", ["enable_flash", "enable_math", "enable_mem_efficient"]) @@ -155,16 +150,6 @@ def Sequential(*mods): return nn.Sequential(*filter(exists, mods)) -def exists(x): - return x is not None - - -def default(val, d): - if exists(val): - return val - return d() if callable(d) else d - - class RMSNorm(nn.Module): def __init__(self, dim, scale=True, dim_cond=None): super().__init__() @@ -202,12 +187,6 @@ def forward(self, x): return super().forward(causal_padded_x) -class GEGLU(nn.Module): - def forward(self, x): - x, gate = x.chunk(2, dim=-1) - return F.gelu(gate) * x - - def FeedForward(dim, mult=4, causal_conv=False): dim_inner = int(dim * mult * 2 / 3) diff --git a/TTS/tts/layers/xtts/stream_generator.py b/TTS/tts/layers/xtts/stream_generator.py index e12f8995cf..44cf940c69 100644 --- a/TTS/tts/layers/xtts/stream_generator.py +++ b/TTS/tts/layers/xtts/stream_generator.py @@ -4,7 +4,7 @@ import inspect import random import warnings -from typing import Callable, List, Optional, Union +from typing import Callable, Optional, Union import numpy as np import torch @@ -20,11 +20,14 @@ PhrasalConstraint, PreTrainedModel, StoppingCriteriaList, + TemperatureLogitsWarper, + TopKLogitsWarper, + TopPLogitsWarper, ) from transformers.generation.utils import GenerateOutput, SampleOutput, logger -def setup_seed(seed): +def setup_seed(seed: int) -> None: if seed == -1: return torch.manual_seed(seed) @@ -43,15 +46,15 @@ def __init__(self, **kwargs): class NewGenerationMixin(GenerationMixin): @torch.no_grad() - def generate( + def generate( # noqa: PLR0911 self, inputs: Optional[torch.Tensor] = None, generation_config: Optional[StreamGenerationConfig] = None, logits_processor: Optional[LogitsProcessorList] = None, stopping_criteria: Optional[StoppingCriteriaList] = None, - prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None, + prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], list[int]]] = None, synced_gpus: Optional[bool] = False, - seed=0, + seed: int = 0, **kwargs, ) -> Union[GenerateOutput, torch.LongTensor]: r""" @@ -90,7 +93,7 @@ def generate( Custom stopping criteria that complement the default stopping criteria built from arguments and a generation config. If a stopping criteria is passed that is already created with the arguments or a generation config an error is thrown. This feature is intended for advanced users. - prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], List[int]]`, *optional*): + prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], list[int]]`, *optional*): If provided, this function constraints the beam search to allowed tokens only at each step. If not provided no constraint is applied. This function takes 2 arguments: the batch ID `batch_id` and `input_ids`. It has to return a list with the allowed tokens for the next generation step conditioned @@ -178,15 +181,34 @@ def generate( model_kwargs["output_attentions"] = generation_config.output_attentions model_kwargs["output_hidden_states"] = generation_config.output_hidden_states model_kwargs["use_cache"] = generation_config.use_cache + model_kwargs["cache_position"] = torch.Tensor([0]).to(inputs_tensor.device) accepts_attention_mask = "attention_mask" in set(inspect.signature(self.forward).parameters.keys()) requires_attention_mask = "encoder_outputs" not in model_kwargs if model_kwargs.get("attention_mask", None) is None and requires_attention_mask and accepts_attention_mask: + setattr( + generation_config, + "_pad_token_tensor", + torch.full( + (inputs_tensor.shape[0], inputs_tensor.shape[1]), + generation_config.pad_token_id, + device=inputs_tensor.device, + ), + ) + setattr( + generation_config, + "_eos_token_tensor", + torch.full( + (inputs_tensor.shape[0], inputs_tensor.shape[1]), + generation_config.eos_token_id, + device=inputs_tensor.device, + ), + ) model_kwargs["attention_mask"] = self._prepare_attention_mask_for_generation( inputs_tensor, - generation_config.pad_token_id, - generation_config.eos_token_id, + generation_config._pad_token_tensor, + generation_config._eos_token_tensor, ) # decoder-only models should use left-padding for generation @@ -384,7 +406,7 @@ def generate( elif is_sample_gen_mode: # 11. prepare logits warper - logits_warper = self._get_logits_warper(generation_config) + logits_warper = _get_logits_warper(generation_config) # 12. expand input_ids with `num_return_sequences` additional sequences per batch input_ids, model_kwargs = self._expand_inputs_for_generation( @@ -409,7 +431,7 @@ def generate( ) elif is_sample_gen_stream_mode: # 11. prepare logits warper - logits_warper = self._get_logits_warper(generation_config) + logits_warper = _get_logits_warper(generation_config) # 12. expand input_ids with `num_return_sequences` additional sequences per batch input_ids, model_kwargs = self._expand_inputs_for_generation( @@ -471,7 +493,7 @@ def generate( elif is_beam_sample_gen_mode: # 11. prepare logits warper - logits_warper = self._get_logits_warper(generation_config) + logits_warper = _get_logits_warper(generation_config) if stopping_criteria.max_length is None: raise ValueError("`max_length` needs to be a stopping_criteria for now.") @@ -577,7 +599,7 @@ def generate( def typeerror(): raise ValueError( - "`force_words_ids` has to either be a `List[List[List[int]]]` or `List[List[int]]`" + "`force_words_ids` has to either be a `list[list[list[int]]]` or `list[list[int]]`" f"of positive integers, but is {generation_config.force_words_ids}." ) @@ -649,7 +671,7 @@ def sample_stream( logits_warper: Optional[LogitsProcessorList] = None, max_length: Optional[int] = None, pad_token_id: Optional[int] = None, - eos_token_id: Optional[Union[int, List[int]]] = None, + eos_token_id: Optional[Union[int, list[int]]] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_scores: Optional[bool] = None, @@ -928,3 +950,17 @@ def init_stream_support(): chunk = tokenizer.decode(x, skip_special_tokens=True) stream_result += chunk print(stream_result) + + +def _get_logits_warper(generation_config: GenerationConfig) -> LogitsProcessorList: + + warpers = LogitsProcessorList() + + if generation_config.temperature is not None and generation_config.temperature != 1.0: + warpers.append(TemperatureLogitsWarper(generation_config.temperature)) + if generation_config.top_k is not None and generation_config.top_k != 0: + warpers.append(TopKLogitsWarper(top_k=generation_config.top_k, min_tokens_to_keep=1)) + if generation_config.top_p is not None and generation_config.top_p < 1.0: + warpers.append(TopPLogitsWarper(top_p=generation_config.top_p, min_tokens_to_keep=1)) + + return warpers diff --git a/TTS/tts/layers/xtts/tokenizer.py b/TTS/tts/layers/xtts/tokenizer.py index 1a3cc47aaf..fec8358deb 100644 --- a/TTS/tts/layers/xtts/tokenizer.py +++ b/TTS/tts/layers/xtts/tokenizer.py @@ -1,24 +1,27 @@ +import logging import os import re import textwrap from functools import cached_property -import pypinyin import torch -from hangul_romanize import Transliter -from hangul_romanize.rule import academic from num2words import num2words from spacy.lang.ar import Arabic from spacy.lang.en import English from spacy.lang.es import Spanish +from spacy.lang.hi import Hindi from spacy.lang.ja import Japanese from spacy.lang.zh import Chinese from tokenizers import Tokenizer from TTS.tts.layers.xtts.zh_num2words import TextNorm as zh_num2words +from TTS.tts.utils.text.cleaners import collapse_whitespace, lowercase + +logger = logging.getLogger(__name__) def get_spacy_lang(lang): + """Return Spacy language used for sentence splitting.""" if lang == "zh": return Chinese() elif lang == "ja": @@ -27,8 +30,10 @@ def get_spacy_lang(lang): return Arabic() elif lang == "es": return Spanish() + elif lang == "hi": + return Hindi() else: - # For most languages, Enlish does the job + # For most languages, English does the job return English() @@ -68,8 +73,6 @@ def split_sentence(text, lang, text_split_length=250): return text_splits -_whitespace_re = re.compile(r"\s+") - # List of (regular expression, replacement) pairs for abbreviations: _abbreviations = { "en": [ @@ -229,6 +232,12 @@ def split_sentence(text, lang, text_split_length=250): # Korean doesn't typically use abbreviations in the same way as Latin-based scripts. ] ], + "hi": [ + (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1]) + for x in [ + # Hindi doesn't typically use abbreviations in the same way as Latin-based scripts. + ] + ], } @@ -425,6 +434,18 @@ def expand_abbreviations_multilingual(text, lang="en"): ("°", " 도 "), ] ], + "hi": [ + (re.compile(r"%s" % re.escape(x[0]), re.IGNORECASE), x[1]) + for x in [ + ("&", " ā¤”ā¤° "), + ("@", " ā¤ā¤Ÿ ā¤ĻāĨ€ ā¤°āĨ‡ā¤Ÿ "), + ("%", " ā¤ĒāĨā¤°ā¤¤ā¤ŋā¤ļā¤¤ "), + ("#", " ā¤šāĨˆā¤ļ "), + ("$", " ā¤ĄāĨ‰ā¤˛ā¤° "), + ("ÂŖ", " ā¤Ēā¤žā¤‰ā¤‚ā¤Ą "), + ("°", " ā¤Ąā¤ŋā¤—āĨā¤°āĨ€ "), + ] + ], } @@ -450,6 +471,7 @@ def expand_symbols_multilingual(text, lang="en"): "tr": re.compile(r"([0-9]+)(\.|inci|nci|uncu|ÃŧncÃŧ|\.)"), "hu": re.compile(r"([0-9]+)(\.|adik|edik|odik|edik|Ãļdik|Ãļdike|ik)"), "ko": re.compile(r"([0-9]+)(번ė§¸|번|ė°¨|ė§¸)"), + "hi": re.compile(r"([0-9]+)(st|nd|rd|th)"), # To check } _number_re = re.compile(r"[0-9]+") _currency_re = { @@ -479,12 +501,12 @@ def _remove_dots(m): def _expand_decimal_point(m, lang="en"): amount = m.group(1).replace(",", ".") - return num2words(float(amount), lang=lang if lang != "cs" else "cz") + return num2words(float(amount), lang=lang) def _expand_currency(m, lang="en", currency="USD"): amount = float((re.sub(r"[^\d.]", "", m.group(0).replace(",", ".")))) - full_amount = num2words(amount, to="currency", currency=currency, lang=lang if lang != "cs" else "cz") + full_amount = num2words(amount, to="currency", currency=currency, lang=lang) and_equivalents = { "en": ", ", @@ -501,6 +523,7 @@ def _expand_currency(m, lang="en", currency="USD"): "tr": ", ", "hu": ", ", "ko": ", ", + "hi": ", ", } if amount.is_integer(): @@ -512,11 +535,11 @@ def _expand_currency(m, lang="en", currency="USD"): def _expand_ordinal(m, lang="en"): - return num2words(int(m.group(1)), ordinal=True, lang=lang if lang != "cs" else "cz") + return num2words(int(m.group(1)), ordinal=True, lang=lang) def _expand_number(m, lang="en"): - return num2words(int(m.group(0)), lang=lang if lang != "cs" else "cz") + return num2words(int(m.group(0)), lang=lang) def expand_numbers_multilingual(text, lang="en"): @@ -540,14 +563,6 @@ def expand_numbers_multilingual(text, lang="en"): return text -def lowercase(text): - return text.lower() - - -def collapse_whitespace(text): - return re.sub(_whitespace_re, " ", text) - - def multilingual_cleaners(text, lang): text = text.replace('"', "") if lang == "tr": @@ -562,14 +577,11 @@ def multilingual_cleaners(text, lang): return text -def basic_cleaners(text): - """Basic pipeline that lowercases and collapses whitespace without transliteration.""" - text = lowercase(text) - text = collapse_whitespace(text) - return text - - def chinese_transliterate(text): + try: + import pypinyin + except ImportError as e: + raise ImportError("Chinese requires: pypinyin") from e return "".join( [p[0] for p in pypinyin.pinyin(text, style=pypinyin.Style.TONE3, heteronym=False, neutral_tone_with_five=True)] ) @@ -582,6 +594,11 @@ def japanese_cleaners(text, katsu): def korean_transliterate(text): + try: + from hangul_romanize import Transliter + from hangul_romanize.rule import academic + except ImportError as e: + raise ImportError("Korean requires: hangul_romanize") from e r = Transliter(academic) return r.translit(text) @@ -611,6 +628,7 @@ def __init__(self, vocab_file=None): "ja": 71, "hu": 224, "ko": 95, + "hi": 150, } @cached_property @@ -623,12 +641,14 @@ def check_input_length(self, txt, lang): lang = lang.split("-")[0] # remove the region limit = self.char_limits.get(lang, 250) if len(txt) > limit: - print( - f"[!] Warning: The text length exceeds the character limit of {limit} for language '{lang}', this might cause truncated audio." + logger.warning( + "The text length exceeds the character limit of %d for language '%s', this might cause truncated audio.", + limit, + lang, ) def preprocess_text(self, txt, lang): - if lang in {"ar", "cs", "de", "en", "es", "fr", "hu", "it", "nl", "pl", "pt", "ru", "tr", "zh", "ko"}: + if lang in {"ar", "cs", "de", "en", "es", "fr", "hi", "hu", "it", "nl", "pl", "pt", "ru", "tr", "zh", "ko"}: txt = multilingual_cleaners(txt, lang) if lang == "zh": txt = chinese_transliterate(txt) @@ -636,9 +656,6 @@ def preprocess_text(self, txt, lang): txt = korean_transliterate(txt) elif lang == "ja": txt = japanese_cleaners(txt, self.katsu) - elif lang == "hi": - # @manmay will implement this - txt = basic_cleaners(txt) else: raise NotImplementedError(f"Language '{lang}' is not supported.") return txt @@ -761,6 +778,9 @@ def test_expand_numbers_multilingual(): ("12.5 ė´ˆ ė•ˆė—.", "ė‹­ė´ ė  다ė„¯ ė´ˆ ė•ˆė—.", "ko"), ("50 ëĒ…ė˜ ëŗ‘ė‚Ŧ가 ėžˆė—ˆë‹¤.", "ė˜¤ė‹­ ëĒ…ė˜ ëŗ‘ė‚Ŧ가 ėžˆė—ˆë‹¤.", "ko"), ("ė´ę˛ƒė€ 1 번ė§¸ 테ėŠ¤íŠ¸ėž…니다", "ė´ę˛ƒė€ ė˛Ģ 번ė§¸ 테ėŠ¤íŠ¸ėž…니다", "ko"), + # Hindi + ("12.5 ā¤¸āĨ‡ā¤•ā¤‚ā¤Ą ā¤ŽāĨ‡ā¤‚āĨ¤", "ā¤¸ā¤žāĨāĨ‡ ā¤Ŧā¤žā¤°ā¤š ā¤¸āĨ‡ā¤•ā¤‚ā¤Ą ā¤ŽāĨ‡ā¤‚āĨ¤", "hi"), + ("50 ā¤¸āĨˆā¤¨ā¤ŋā¤• ā¤ĨāĨ‡āĨ¤", "ā¤Ēā¤šā¤žā¤¸ ā¤¸āĨˆā¤¨ā¤ŋā¤• ā¤ĨāĨ‡āĨ¤", "hi"), ] for a, b, lang in test_cases: out = expand_numbers_multilingual(a, lang=lang) @@ -830,6 +850,7 @@ def test_symbols_multilingual(): ("Pilim %14 dolu.", "Pilim yÃŧzde 14 dolu.", "tr"), ("Az akkumulÃĄtorom tÃļltÃļttsÊge 14%", "Az akkumulÃĄtorom tÃļltÃļttsÊge 14 szÃĄzalÊk", "hu"), ("배터ëĻŦ ėž”량ė´ 14%ėž…니다.", "배터ëĻŦ ėž”량ė´ 14 íŧė„ŧ트ėž…니다.", "ko"), + ("ā¤ŽāĨ‡ā¤°āĨ‡ ā¤Ēā¤žā¤¸ 14% ā¤ŦāĨˆā¤Ÿā¤°āĨ€ ā¤šāĨˆāĨ¤", "ā¤ŽāĨ‡ā¤°āĨ‡ ā¤Ēā¤žā¤¸ ā¤šāĨŒā¤Ļā¤š ā¤ĒāĨā¤°ā¤¤ā¤ŋā¤ļā¤¤ ā¤ŦāĨˆā¤Ÿā¤°āĨ€ ā¤šāĨˆāĨ¤", "hi"), ] for a, b, lang in test_cases: diff --git a/TTS/tts/layers/xtts/trainer/dataset.py b/TTS/tts/layers/xtts/trainer/dataset.py index 2f958cb5a5..e598232665 100644 --- a/TTS/tts/layers/xtts/trainer/dataset.py +++ b/TTS/tts/layers/xtts/trainer/dataset.py @@ -1,4 +1,4 @@ -import os +import logging import random import sys @@ -8,6 +8,8 @@ from TTS.tts.models.xtts import load_audio +logger = logging.getLogger(__name__) + torch.set_num_threads(1) @@ -71,13 +73,13 @@ def __init__(self, config, samples, tokenizer, sample_rate, is_eval=False): random.shuffle(self.samples) # order by language self.samples = key_samples_by_col(self.samples, "language") - print(" > Sampling by language:", self.samples.keys()) + logger.info("Sampling by language: %s", self.samples.keys()) else: # for evaluation load and check samples that are corrupted to ensures the reproducibility self.check_eval_samples() def check_eval_samples(self): - print(" > Filtering invalid eval samples!!") + logger.info("Filtering invalid eval samples!!") new_samples = [] for sample in self.samples: try: @@ -93,7 +95,7 @@ def check_eval_samples(self): continue new_samples.append(sample) self.samples = new_samples - print(" > Total eval samples after filtering:", len(self.samples)) + logger.info("Total eval samples after filtering: %d", len(self.samples)) def get_text(self, text, lang): tokens = self.tokenizer.encode(text, lang) @@ -151,7 +153,7 @@ def __getitem__(self, index): # ignore samples that we already know that is not valid ones if sample_id in self.failed_samples: if self.debug_failures: - print(f"Ignoring sample {sample['audio_file']} because it was already ignored before !!") + logger.info("Ignoring sample %s because it was already ignored before !!", sample["audio_file"]) # call get item again to get other sample return self[1] @@ -160,7 +162,7 @@ def __getitem__(self, index): tseq, audiopath, wav, cond, cond_len, cond_idxs = self.load_item(sample) except: if self.debug_failures: - print(f"error loading {sample['audio_file']} {sys.exc_info()}") + logger.warning("Error loading %s %s", sample["audio_file"], sys.exc_info()) self.failed_samples.add(sample_id) return self[1] @@ -173,8 +175,11 @@ def __getitem__(self, index): # Basically, this audio file is nonexistent or too long to be supported by the dataset. # It's hard to handle this situation properly. Best bet is to return the a random valid token and skew the dataset somewhat as a result. if self.debug_failures and wav is not None and tseq is not None: - print( - f"error loading {sample['audio_file']}: ranges are out of bounds; {wav.shape[-1]}, {tseq.shape[0]}" + logger.warning( + "Error loading %s: ranges are out of bounds: %d, %d", + sample["audio_file"], + wav.shape[-1], + tseq.shape[0], ) self.failed_samples.add(sample_id) return self[1] @@ -187,9 +192,9 @@ def __getitem__(self, index): "wav_lengths": torch.tensor(wav.shape[-1], dtype=torch.long), "filenames": audiopath, "conditioning": cond.unsqueeze(1), - "cond_lens": torch.tensor(cond_len, dtype=torch.long) - if cond_len is not torch.nan - else torch.tensor([cond_len]), + "cond_lens": ( + torch.tensor(cond_len, dtype=torch.long) if cond_len is not torch.nan else torch.tensor([cond_len]) + ), "cond_idxs": torch.tensor(cond_idxs) if cond_idxs is not torch.nan else torch.tensor([cond_idxs]), } return res diff --git a/TTS/tts/layers/xtts/trainer/gpt_trainer.py b/TTS/tts/layers/xtts/trainer/gpt_trainer.py index 9a7a1d7783..107054189c 100644 --- a/TTS/tts/layers/xtts/trainer/gpt_trainer.py +++ b/TTS/tts/layers/xtts/trainer/gpt_trainer.py @@ -1,3 +1,4 @@ +import logging from dataclasses import dataclass, field from typing import Dict, List, Tuple, Union @@ -5,8 +6,8 @@ import torch.nn as nn import torchaudio from coqpit import Coqpit -from torch.nn import functional as F from torch.utils.data import DataLoader +from trainer.io import load_fsspec from trainer.torch import DistributedSampler from trainer.trainer_utils import get_optimizer, get_scheduler @@ -17,8 +18,10 @@ from TTS.tts.layers.xtts.tokenizer import VoiceBpeTokenizer from TTS.tts.layers.xtts.trainer.dataset import XTTSDataset from TTS.tts.models.base_tts import BaseTTS -from TTS.tts.models.xtts import Xtts, XttsArgs, XttsAudioConfig -from TTS.utils.io import load_fsspec +from TTS.tts.models.xtts import Xtts, XttsArgs +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 + +logger = logging.getLogger(__name__) @dataclass @@ -31,11 +34,6 @@ class GPTTrainerConfig(XttsConfig): test_sentences: List[dict] = field(default_factory=lambda: []) -@dataclass -class XttsAudioConfig(XttsAudioConfig): - dvae_sample_rate: int = 22050 - - @dataclass class GPTArgs(XttsArgs): min_conditioning_length: int = 66150 @@ -47,7 +45,7 @@ class GPTArgs(XttsArgs): max_wav_length: int = 255995 # ~11.6 seconds max_text_length: int = 200 tokenizer_file: str = "" - mel_norm_file: str = "https://coqui.gateway.scarf.sh/v0.14.0_models/mel_norms.pth" + mel_norm_file: str = "https://github.com/coqui-ai/TTS/releases/download/v0.14.0_models/mel_norms.pth" dvae_checkpoint: str = "" xtts_checkpoint: str = "" gpt_checkpoint: str = "" # if defined it will replace the gpt weights on xtts model @@ -58,7 +56,7 @@ def callback_clearml_load_save(operation_type, model_info): # return None means skip the file upload/log, returning model_info will continue with the log/upload # you can also change the upload destination file name model_info.upload_filename or check the local file size with Path(model_info.local_model_path).stat().st_size assert operation_type in ("load", "save") - # print(operation_type, model_info.__dict__) + logger.debug("%s %s", operation_type, model_info.__dict__) if "similarities.pth" in model_info.__dict__["local_model_path"]: return None @@ -89,10 +87,12 @@ def __init__(self, config: Coqpit): # load GPT if available if self.args.gpt_checkpoint: - gpt_checkpoint = torch.load(self.args.gpt_checkpoint, map_location=torch.device("cpu")) + gpt_checkpoint = torch.load( + self.args.gpt_checkpoint, map_location=torch.device("cpu"), weights_only=is_pytorch_at_least_2_4() + ) # deal with coqui Trainer exported model if "model" in gpt_checkpoint.keys() and "config" in gpt_checkpoint.keys(): - print("Coqui Trainer checkpoint detected! Converting it!") + logger.info("Coqui Trainer checkpoint detected! Converting it!") gpt_checkpoint = gpt_checkpoint["model"] states_keys = list(gpt_checkpoint.keys()) for key in states_keys: @@ -111,7 +111,7 @@ def __init__(self, config: Coqpit): num_new_tokens = ( self.xtts.gpt.text_embedding.weight.shape[0] - gpt_checkpoint["text_embedding.weight"].shape[0] ) - print(f" > Loading checkpoint with {num_new_tokens} additional tokens.") + logger.info("Loading checkpoint with %d additional tokens.", num_new_tokens) # add new tokens to a linear layer (text_head) emb_g = gpt_checkpoint["text_embedding.weight"] @@ -138,7 +138,7 @@ def __init__(self, config: Coqpit): gpt_checkpoint["text_head.bias"] = text_head_bias self.xtts.gpt.load_state_dict(gpt_checkpoint, strict=True) - print(">> GPT weights restored from:", self.args.gpt_checkpoint) + logger.info("GPT weights restored from: %s", self.args.gpt_checkpoint) # Mel spectrogram extractor for conditioning if self.args.gpt_use_perceiver_resampler: @@ -182,9 +182,11 @@ def __init__(self, config: Coqpit): self.dvae.eval() if self.args.dvae_checkpoint: - dvae_checkpoint = torch.load(self.args.dvae_checkpoint, map_location=torch.device("cpu")) + dvae_checkpoint = torch.load( + self.args.dvae_checkpoint, map_location=torch.device("cpu"), weights_only=is_pytorch_at_least_2_4() + ) self.dvae.load_state_dict(dvae_checkpoint, strict=False) - print(">> DVAE weights restored from:", self.args.dvae_checkpoint) + logger.info("DVAE weights restored from: %s", self.args.dvae_checkpoint) else: raise RuntimeError( "You need to specify config.model_args.dvae_checkpoint path to be able to train the GPT decoder!!" @@ -230,7 +232,7 @@ def test_run(self, assets) -> Tuple[Dict, Dict]: # pylint: disable=W0613 # init gpt for inference mode self.xtts.gpt.init_gpt_for_inference(kv_cache=self.args.kv_cache, use_deepspeed=False) self.xtts.gpt.eval() - print(" | > Synthesizing test sentences.") + logger.info("Synthesizing test sentences.") for idx, s_info in enumerate(self.config.test_sentences): wav = self.xtts.synthesize( s_info["text"], @@ -391,7 +393,7 @@ def get_data_loader( loader = DataLoader( dataset, sampler=sampler, - batch_size = config.eval_batch_size if is_eval else config.batch_size, + batch_size=config.eval_batch_size if is_eval else config.batch_size, collate_fn=dataset.collate_fn, num_workers=config.num_eval_loader_workers if is_eval else config.num_loader_workers, pin_memory=False, diff --git a/TTS/tts/layers/xtts/xtts_manager.py b/TTS/tts/layers/xtts/xtts_manager.py index 3e7d0f6c91..8156b35f0d 100644 --- a/TTS/tts/layers/xtts/xtts_manager.py +++ b/TTS/tts/layers/xtts/xtts_manager.py @@ -1,34 +1,37 @@ import torch -class SpeakerManager(): +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 + + +class SpeakerManager: def __init__(self, speaker_file_path=None): - self.speakers = torch.load(speaker_file_path) + self.speakers = torch.load(speaker_file_path, weights_only=is_pytorch_at_least_2_4()) @property def name_to_id(self): - return self.speakers.keys() - + return self.speakers + @property def num_speakers(self): return len(self.name_to_id) - + @property def speaker_names(self): return list(self.name_to_id.keys()) - -class LanguageManager(): + +class LanguageManager: def __init__(self, config): self.langs = config["languages"] @property def name_to_id(self): return self.langs - + @property def num_languages(self): return len(self.name_to_id) - + @property def language_names(self): return list(self.name_to_id) diff --git a/TTS/tts/layers/xtts/zh_num2words.py b/TTS/tts/layers/xtts/zh_num2words.py index e59ccb6630..69b8dae952 100644 --- a/TTS/tts/layers/xtts/zh_num2words.py +++ b/TTS/tts/layers/xtts/zh_num2words.py @@ -4,13 +4,14 @@ import argparse import csv -import os +import logging import re import string import sys -# fmt: off +logger = logging.getLogger(__name__) +# fmt: off # ================================================================================ # # basic constant # ================================================================================ # @@ -491,8 +492,6 @@ class NumberSystem(object): 中文数字įŗģįģŸ """ - pass - class MathSymbol(object): """ @@ -927,12 +926,13 @@ def percentage2chntext(self): def normalize_nsw(raw_text): text = "^" + raw_text + "$" + logger.debug(text) # č§„čŒƒåŒ–æ—Ĩ期 pattern = re.compile(r"\D+((([089]\d|(19|20)\d{2})åš´)?(\d{1,2}月(\d{1,2}[æ—Ĩåˇ])?)?)") matchers = pattern.findall(text) if matchers: - # print('date') + logger.debug("date") for matcher in matchers: text = text.replace(matcher[0], Date(date=matcher[0]).date2chntext(), 1) @@ -940,7 +940,7 @@ def normalize_nsw(raw_text): pattern = re.compile(r"\D+((\d+(\.\d+)?)[多äŊ™å‡ ]?" + CURRENCY_UNITS + r"(\d" + CURRENCY_UNITS + r"?)?)") matchers = pattern.findall(text) if matchers: - # print('money') + logger.debug("money") for matcher in matchers: text = text.replace(matcher[0], Money(money=matcher[0]).money2chntext(), 1) @@ -953,14 +953,14 @@ def normalize_nsw(raw_text): pattern = re.compile(r"\D((\+?86 ?)?1([38]\d|5[0-35-9]|7[678]|9[89])\d{8})\D") matchers = pattern.findall(text) if matchers: - # print('telephone') + logger.debug("telephone") for matcher in matchers: text = text.replace(matcher[0], TelePhone(telephone=matcher[0]).telephone2chntext(), 1) # å›ēč¯ pattern = re.compile(r"\D((0(10|2[1-3]|[3-9]\d{2})-?)?[1-9]\d{6,7})\D") matchers = pattern.findall(text) if matchers: - # print('fixed telephone') + logger.debug("fixed telephone") for matcher in matchers: text = text.replace(matcher[0], TelePhone(telephone=matcher[0]).telephone2chntext(fixed=True), 1) @@ -968,7 +968,7 @@ def normalize_nsw(raw_text): pattern = re.compile(r"(\d+/\d+)") matchers = pattern.findall(text) if matchers: - # print('fraction') + logger.debug("fraction") for matcher in matchers: text = text.replace(matcher, Fraction(fraction=matcher).fraction2chntext(), 1) @@ -977,7 +977,7 @@ def normalize_nsw(raw_text): pattern = re.compile(r"(\d+(\.\d+)?%)") matchers = pattern.findall(text) if matchers: - # print('percentage') + logger.debug("percentage") for matcher in matchers: text = text.replace(matcher[0], Percentage(percentage=matcher[0]).percentage2chntext(), 1) @@ -985,7 +985,7 @@ def normalize_nsw(raw_text): pattern = re.compile(r"(\d+(\.\d+)?)[多äŊ™å‡ ]?" + COM_QUANTIFIERS) matchers = pattern.findall(text) if matchers: - # print('cardinal+quantifier') + logger.debug("cardinal+quantifier") for matcher in matchers: text = text.replace(matcher[0], Cardinal(cardinal=matcher[0]).cardinal2chntext(), 1) @@ -993,7 +993,7 @@ def normalize_nsw(raw_text): pattern = re.compile(r"(\d{4,32})") matchers = pattern.findall(text) if matchers: - # print('digit') + logger.debug("digit") for matcher in matchers: text = text.replace(matcher, Digit(digit=matcher).digit2chntext(), 1) @@ -1001,7 +1001,7 @@ def normalize_nsw(raw_text): pattern = re.compile(r"(\d+(\.\d+)?)") matchers = pattern.findall(text) if matchers: - # print('cardinal') + logger.debug("cardinal") for matcher in matchers: text = text.replace(matcher[0], Cardinal(cardinal=matcher[0]).cardinal2chntext(), 1) @@ -1009,7 +1009,7 @@ def normalize_nsw(raw_text): pattern = re.compile(r"(([a-zA-Z]+)äēŒ([a-zA-Z]+))") matchers = pattern.findall(text) if matchers: - # print('particular') + logger.debug("particular") for matcher in matchers: text = text.replace(matcher[0], matcher[1] + "2" + matcher[2], 1) @@ -1107,7 +1107,7 @@ def __call__(self, text): if self.check_chars: for c in text: if not IN_VALID_CHARS.get(c): - print(f"WARNING: illegal char {c} in: {text}", file=sys.stderr) + logger.warning("Illegal char %s in: %s", c, text) return "" if self.remove_space: diff --git a/TTS/tts/models/__init__.py b/TTS/tts/models/__init__.py index 2bd2e5f087..ebfa171c80 100644 --- a/TTS/tts/models/__init__.py +++ b/TTS/tts/models/__init__.py @@ -1,10 +1,13 @@ +import logging from typing import Dict, List, Union from TTS.utils.generic_utils import find_module +logger = logging.getLogger(__name__) + def setup_model(config: "Coqpit", samples: Union[List[List], List[Dict]] = None) -> "BaseTTS": - print(" > Using model: {}".format(config.model)) + logger.info("Using model: %s", config.model) # fetch the right model implementation. if "base_model" in config and config["base_model"] is not None: MyModel = find_module("TTS.tts.models", config.base_model.lower()) diff --git a/TTS/tts/models/align_tts.py b/TTS/tts/models/align_tts.py index b2e51de7d6..28a52bc558 100644 --- a/TTS/tts/models/align_tts.py +++ b/TTS/tts/models/align_tts.py @@ -3,7 +3,9 @@ import torch from coqpit import Coqpit +from monotonic_alignment_search import maximum_path from torch import nn +from trainer.io import load_fsspec from TTS.tts.layers.align_tts.mdn import MDNBlock from TTS.tts.layers.feed_forward.decoder import Decoder @@ -11,11 +13,10 @@ from TTS.tts.layers.feed_forward.encoder import Encoder from TTS.tts.layers.generic.pos_encoding import PositionalEncoding from TTS.tts.models.base_tts import BaseTTS -from TTS.tts.utils.helpers import generate_path, maximum_path, sequence_mask +from TTS.tts.utils.helpers import expand_encoder_outputs, generate_attention, sequence_mask from TTS.tts.utils.speakers import SpeakerManager from TTS.tts.utils.text.tokenizer import TTSTokenizer from TTS.tts.utils.visual import plot_alignment, plot_spectrogram -from TTS.utils.io import load_fsspec @dataclass @@ -168,35 +169,6 @@ def compute_align_path(self, mu, log_sigma, y, x_mask, y_mask): dr_mas = torch.sum(attn, -1) return dr_mas.squeeze(1), log_p - @staticmethod - def generate_attn(dr, x_mask, y_mask=None): - # compute decode mask from the durations - if y_mask is None: - y_lengths = dr.sum(1).long() - y_lengths[y_lengths < 1] = 1 - y_mask = torch.unsqueeze(sequence_mask(y_lengths, None), 1).to(dr.dtype) - attn_mask = torch.unsqueeze(x_mask, -1) * torch.unsqueeze(y_mask, 2) - attn = generate_path(dr, attn_mask.squeeze(1)).to(dr.dtype) - return attn - - def expand_encoder_outputs(self, en, dr, x_mask, y_mask): - """Generate attention alignment map from durations and - expand encoder outputs - - Examples:: - - encoder output: [a,b,c,d] - - durations: [1, 3, 2, 1] - - - expanded: [a, b, b, b, c, c, d] - - attention map: [[0, 0, 0, 0, 0, 0, 1], - [0, 0, 0, 0, 1, 1, 0], - [0, 1, 1, 1, 0, 0, 0], - [1, 0, 0, 0, 0, 0, 0]] - """ - attn = self.generate_attn(dr, x_mask, y_mask) - o_en_ex = torch.matmul(attn.squeeze(1).transpose(1, 2), en.transpose(1, 2)).transpose(1, 2) - return o_en_ex, attn - def format_durations(self, o_dr_log, x_mask): o_dr = (torch.exp(o_dr_log) - 1) * x_mask * self.length_scale o_dr[o_dr < 1] = 1.0 @@ -242,9 +214,8 @@ def _forward_encoder(self, x, x_lengths, g=None): return o_en, o_en_dp, x_mask, g def _forward_decoder(self, o_en, o_en_dp, dr, x_mask, y_lengths, g): - y_mask = torch.unsqueeze(sequence_mask(y_lengths, None), 1).to(o_en_dp.dtype) # expand o_en with durations - o_en_ex, attn = self.expand_encoder_outputs(o_en, dr, x_mask, y_mask) + o_en_ex, attn, y_mask = expand_encoder_outputs(o_en, dr, x_mask, y_lengths) # positional encoding if hasattr(self, "pos_encoder"): o_en_ex = self.pos_encoder(o_en_ex, y_mask) @@ -281,7 +252,7 @@ def forward( o_en, o_en_dp, x_mask, g = self._forward_encoder(x, x_lengths, g) dr_mas, mu, log_sigma, logp = self._forward_mdn(o_en, y, y_lengths, x_mask) y_mask = torch.unsqueeze(sequence_mask(y_lengths, None), 1).to(o_en_dp.dtype) - attn = self.generate_attn(dr_mas, x_mask, y_mask) + attn = generate_attention(dr_mas, x_mask, y_mask) elif phase == 1: # train decoder o_en, o_en_dp, x_mask, g = self._forward_encoder(x, x_lengths, g) @@ -415,7 +386,7 @@ def _set_phase(config, global_step): """Decide AlignTTS training phase""" if isinstance(config.phase_start_steps, list): vals = [i < global_step for i in config.phase_start_steps] - if not True in vals: + if True not in vals: phase = 0 else: phase = ( diff --git a/TTS/tts/models/bark.py b/TTS/tts/models/bark.py index e5edffd4ef..c52c541b25 100644 --- a/TTS/tts/models/bark.py +++ b/TTS/tts/models/bark.py @@ -164,7 +164,7 @@ def generate_audio( return audio_arr, [x_semantic, c, f] def generate_voice(self, audio, speaker_id, voice_dir): - """Generate a voice from the given audio and text. + """Generate a voice from the given audio. Args: audio (str): Path to the audio file. @@ -174,7 +174,7 @@ def generate_voice(self, audio, speaker_id, voice_dir): if voice_dir is not None: voice_dirs = [voice_dir] try: - _ = load_voice(speaker_id, voice_dirs) + _ = load_voice(self, speaker_id, voice_dirs) except (KeyError, FileNotFoundError): output_path = os.path.join(voice_dir, speaker_id + ".npz") os.makedirs(voice_dir, exist_ok=True) @@ -206,12 +206,14 @@ def synthesize( speaker_wav (str): Path to the speaker audio file for cloning a new voice. It is cloned and saved in `voice_dirs` with the name `speaker_id`. Defaults to None. voice_dirs (List[str]): List of paths that host reference audio files for speakers. Defaults to None. - **kwargs: Model specific inference settings used by `generate_audio()` and `TTS.tts.layers.bark.inference_funcs.generate_text_semantic(). + **kwargs: Model specific inference settings used by `generate_audio()` and + `TTS.tts.layers.bark.inference_funcs.generate_text_semantic()`. Returns: - A dictionary of the output values with `wav` as output waveform, `deterministic_seed` as seed used at inference, - `text_input` as text token IDs after tokenizer, `voice_samples` as samples used for cloning, `conditioning_latents` - as latents used at inference. + A dictionary of the output values with `wav` as output waveform, + `deterministic_seed` as seed used at inference, `text_input` as text token IDs + after tokenizer, `voice_samples` as samples used for cloning, + `conditioning_latents` as latents used at inference. """ speaker_id = "random" if speaker_id is None else speaker_id @@ -225,14 +227,11 @@ def synthesize( return return_dict - def eval_step(self): - ... + def eval_step(self): ... - def forward(self): - ... + def forward(self): ... - def inference(self): - ... + def inference(self): ... @staticmethod def init_from_config(config: "BarkConfig", **kwargs): # pylint: disable=unused-argument @@ -246,7 +245,6 @@ def load_checkpoint( text_model_path=None, coarse_model_path=None, fine_model_path=None, - hubert_model_path=None, hubert_tokenizer_path=None, eval=False, strict=True, @@ -269,13 +267,11 @@ def load_checkpoint( text_model_path = text_model_path or os.path.join(checkpoint_dir, "text_2.pt") coarse_model_path = coarse_model_path or os.path.join(checkpoint_dir, "coarse_2.pt") fine_model_path = fine_model_path or os.path.join(checkpoint_dir, "fine_2.pt") - hubert_model_path = hubert_model_path or os.path.join(checkpoint_dir, "hubert.pt") hubert_tokenizer_path = hubert_tokenizer_path or os.path.join(checkpoint_dir, "tokenizer.pth") self.config.LOCAL_MODEL_PATHS["text"] = text_model_path self.config.LOCAL_MODEL_PATHS["coarse"] = coarse_model_path self.config.LOCAL_MODEL_PATHS["fine"] = fine_model_path - self.config.LOCAL_MODEL_PATHS["hubert"] = hubert_model_path self.config.LOCAL_MODEL_PATHS["hubert_tokenizer"] = hubert_tokenizer_path self.load_bark_models() diff --git a/TTS/tts/models/base_tacotron.py b/TTS/tts/models/base_tacotron.py index f38dace235..79cdf1a7d4 100644 --- a/TTS/tts/models/base_tacotron.py +++ b/TTS/tts/models/base_tacotron.py @@ -1,10 +1,12 @@ import copy +import logging from abc import abstractmethod from typing import Dict, Tuple import torch from coqpit import Coqpit from torch import nn +from trainer.io import load_fsspec from TTS.tts.layers.losses import TacotronLoss from TTS.tts.models.base_tts import BaseTTS @@ -14,9 +16,10 @@ from TTS.tts.utils.text.tokenizer import TTSTokenizer from TTS.tts.utils.visual import plot_alignment, plot_spectrogram from TTS.utils.generic_utils import format_aux_input -from TTS.utils.io import load_fsspec from TTS.utils.training import gradual_training_scheduler +logger = logging.getLogger(__name__) + class BaseTacotron(BaseTTS): """Base class shared by Tacotron and Tacotron2""" @@ -100,7 +103,8 @@ def load_checkpoint( config (Coqpi): model configuration. checkpoint_path (str): path to checkpoint file. eval (bool, optional): whether to load model for evaluation. - cache (bool, optional): If True, cache the file locally for subsequent calls. It is cached under `get_user_data_dir()/tts_cache`. Defaults to False. + cache (bool, optional): If True, cache the file locally for subsequent calls. + It is cached under `trainer.io.get_user_data_dir()/tts_cache`. Defaults to False. """ state = load_fsspec(checkpoint_path, map_location=torch.device("cpu"), cache=cache) self.load_state_dict(state["model"]) @@ -116,7 +120,7 @@ def load_checkpoint( self.decoder.set_r(config.r) if eval: self.eval() - print(f" > Model's reduction rate `r` is set to: {self.decoder.r}") + logger.info("Model's reduction rate `r` is set to: %d", self.decoder.r) assert not self.training def get_criterion(self) -> nn.Module: @@ -148,7 +152,7 @@ def test_run(self, assets: Dict) -> Tuple[Dict, Dict]: Returns: Tuple[Dict, Dict]: Test figures and audios to be projected to Tensorboard. """ - print(" | > Synthesizing test sentences.") + logger.info("Synthesizing test sentences.") test_audios = {} test_figures = {} test_sentences = self.config.test_sentences @@ -302,4 +306,4 @@ def on_epoch_start(self, trainer): self.decoder.set_r(r) if trainer.config.bidirectional_decoder: trainer.model.decoder_backward.set_r(r) - print(f"\n > Number of output frames: {self.decoder.r}") + logger.info("Number of output frames: %d", self.decoder.r) diff --git a/TTS/tts/models/base_tts.py b/TTS/tts/models/base_tts.py index 371db3c482..91dd6e96d6 100644 --- a/TTS/tts/models/base_tts.py +++ b/TTS/tts/models/base_tts.py @@ -1,3 +1,4 @@ +import logging import os import random from typing import Dict, List, Tuple, Union @@ -15,10 +16,12 @@ from TTS.tts.datasets.dataset import TTSDataset from TTS.tts.utils.data import get_length_balancer_weights from TTS.tts.utils.languages import LanguageManager, get_language_balancer_weights -from TTS.tts.utils.speakers import SpeakerManager, get_speaker_balancer_weights, get_speaker_manager +from TTS.tts.utils.speakers import SpeakerManager, get_speaker_balancer_weights from TTS.tts.utils.synthesis import synthesis from TTS.tts.utils.visual import plot_alignment, plot_spectrogram +logger = logging.getLogger(__name__) + # pylint: skip-file @@ -105,15 +108,17 @@ def adjust_speech_rate(self, gpt_latents, length_scale): return gpt_latents def init_multispeaker(self, config: Coqpit, data: List = None): - """Initialize a speaker embedding layer if needen and define expected embedding channel size for defining - `in_channels` size of the connected layers. + """Set up for multi-speaker TTS. + + Initialize a speaker embedding layer if needed and define expected embedding + channel size for defining `in_channels` size of the connected layers. This implementation yields 3 possible outcomes: - 1. If `config.use_speaker_embedding` and `config.use_d_vector_file are False, do nothing. + 1. If `config.use_speaker_embedding` and `config.use_d_vector_file` are False, do nothing. 2. If `config.use_d_vector_file` is True, set expected embedding channel size to `config.d_vector_dim` or 512. 3. If `config.use_speaker_embedding`, initialize a speaker embedding layer with channel size of - `config.d_vector_dim` or 512. + `config.d_vector_dim` or 512. You can override this function for new models. @@ -133,7 +138,7 @@ def init_multispeaker(self, config: Coqpit, data: List = None): ) # init speaker embedding layer if config.use_speaker_embedding and not config.use_d_vector_file: - print(" > Init speaker_embedding layer.") + logger.info("Init speaker_embedding layer.") self.speaker_embedding = nn.Embedding(self.num_speakers, self.embedded_speaker_dim) self.speaker_embedding.weight.data.normal_(0, 0.3) @@ -169,7 +174,7 @@ def get_aux_input_from_test_sentences(self, sentence_info): if speaker_name is None: d_vector = self.speaker_manager.get_random_embedding() else: - d_vector = self.speaker_manager.get_d_vector_by_name(speaker_name) + d_vector = self.speaker_manager.get_mean_embedding(speaker_name) elif config.use_speaker_embedding: if speaker_name is None: speaker_id = self.speaker_manager.get_random_id() @@ -273,12 +278,12 @@ def get_sampler(self, config: Coqpit, dataset: TTSDataset, num_gpus=1): if getattr(config, "use_language_weighted_sampler", False): alpha = getattr(config, "language_weighted_sampler_alpha", 1.0) - print(" > Using Language weighted sampler with alpha:", alpha) + logger.info("Using Language weighted sampler with alpha: %.2f", alpha) weights = get_language_balancer_weights(data_items) * alpha if getattr(config, "use_speaker_weighted_sampler", False): alpha = getattr(config, "speaker_weighted_sampler_alpha", 1.0) - print(" > Using Speaker weighted sampler with alpha:", alpha) + logger.info("Using Speaker weighted sampler with alpha: %.2f", alpha) if weights is not None: weights += get_speaker_balancer_weights(data_items) * alpha else: @@ -286,7 +291,7 @@ def get_sampler(self, config: Coqpit, dataset: TTSDataset, num_gpus=1): if getattr(config, "use_length_weighted_sampler", False): alpha = getattr(config, "length_weighted_sampler_alpha", 1.0) - print(" > Using Length weighted sampler with alpha:", alpha) + logger.info("Using Length weighted sampler with alpha: %.2f", alpha) if weights is not None: weights += get_length_balancer_weights(data_items) * alpha else: @@ -358,7 +363,6 @@ def get_data_loader( phoneme_cache_path=config.phoneme_cache_path, precompute_num_workers=config.precompute_num_workers, use_noise_augment=False if is_eval else config.use_noise_augment, - verbose=verbose, speaker_id_mapping=speaker_id_mapping, d_vector_mapping=d_vector_mapping if config.use_d_vector_file else None, tokenizer=self.tokenizer, @@ -397,9 +401,11 @@ def _get_test_aux_input( d_vector = (random.sample(sorted(d_vector), 1),) aux_inputs = { - "speaker_id": None - if not self.config.use_speaker_embedding - else random.sample(sorted(self.speaker_manager.name_to_id.values()), 1), + "speaker_id": ( + None + if not self.config.use_speaker_embedding + else random.sample(sorted(self.speaker_manager.name_to_id.values()), 1) + ), "d_vector": d_vector, "style_wav": None, # TODO: handle GST style input } @@ -416,7 +422,7 @@ def test_run(self, assets: Dict) -> Tuple[Dict, Dict]: Returns: Tuple[Dict, Dict]: Test figures and audios to be projected to Tensorboard. """ - print(" | > Synthesizing test sentences.") + logger.info("Synthesizing test sentences.") test_audios = {} test_figures = {} test_sentences = self.config.test_sentences @@ -455,8 +461,8 @@ def on_init_start(self, trainer): if hasattr(trainer.config, "model_args"): trainer.config.model_args.speakers_file = output_path trainer.config.save_json(os.path.join(trainer.output_path, "config.json")) - print(f" > `speakers.pth` is saved to {output_path}.") - print(" > `speakers_file` is updated in the config.json.") + logger.info("`speakers.pth` is saved to: %s", output_path) + logger.info("`speakers_file` is updated in the config.json.") if self.language_manager is not None: output_path = os.path.join(trainer.output_path, "language_ids.json") @@ -465,8 +471,8 @@ def on_init_start(self, trainer): if hasattr(trainer.config, "model_args"): trainer.config.model_args.language_ids_file = output_path trainer.config.save_json(os.path.join(trainer.output_path, "config.json")) - print(f" > `language_ids.json` is saved to {output_path}.") - print(" > `language_ids_file` is updated in the config.json.") + logger.info("`language_ids.json` is saved to: %s", output_path) + logger.info("`language_ids_file` is updated in the config.json.") class BaseTTSE2E(BaseTTS): diff --git a/TTS/tts/models/delightful_tts.py b/TTS/tts/models/delightful_tts.py index b1cf886bea..e6db116081 100644 --- a/TTS/tts/models/delightful_tts.py +++ b/TTS/tts/models/delightful_tts.py @@ -1,3 +1,4 @@ +import logging import os from dataclasses import dataclass, field from itertools import chain @@ -7,316 +8,57 @@ import numpy as np import torch import torch.distributed as dist -import torchaudio from coqpit import Coqpit -from librosa.filters import mel as librosa_mel_fn from torch import nn -from torch.cuda.amp.autocast_mode import autocast -from torch.nn import functional as F from torch.utils.data import DataLoader from torch.utils.data.sampler import WeightedRandomSampler +from trainer.io import load_fsspec from trainer.torch import DistributedSampler, DistributedSamplerWrapper from trainer.trainer_utils import get_optimizer, get_scheduler -from TTS.tts.datasets.dataset import F0Dataset, TTSDataset, _parse_sample +from TTS.tts.datasets.dataset import F0Dataset, TTSDataset, _parse_sample, get_attribute_balancer_weights from TTS.tts.layers.delightful_tts.acoustic_model import AcousticModel -from TTS.tts.layers.losses import ForwardSumLoss, VitsDiscriminatorLoss +from TTS.tts.layers.losses import ( + ForwardSumLoss, + VitsDiscriminatorLoss, + _binary_alignment_loss, + feature_loss, + generator_loss, +) from TTS.tts.layers.vits.discriminator import VitsDiscriminator from TTS.tts.models.base_tts import BaseTTSE2E +from TTS.tts.models.vits import load_audio from TTS.tts.utils.helpers import average_over_durations, compute_attn_prior, rand_segments, segment, sequence_mask from TTS.tts.utils.speakers import SpeakerManager +from TTS.tts.utils.synthesis import embedding_to_torch, id_to_torch, numpy_to_torch from TTS.tts.utils.text.tokenizer import TTSTokenizer from TTS.tts.utils.visual import plot_alignment, plot_avg_pitch, plot_pitch, plot_spectrogram from TTS.utils.audio.numpy_transforms import build_mel_basis, compute_f0 from TTS.utils.audio.numpy_transforms import db_to_amp as db_to_amp_numpy from TTS.utils.audio.numpy_transforms import mel_to_wav as mel_to_wav_numpy from TTS.utils.audio.processor import AudioProcessor -from TTS.utils.io import load_fsspec +from TTS.utils.audio.torch_transforms import wav_to_mel, wav_to_spec from TTS.vocoder.layers.losses import MultiScaleSTFTLoss from TTS.vocoder.models.hifigan_generator import HifiganGenerator from TTS.vocoder.utils.generic_utils import plot_results - -def id_to_torch(aux_id, cuda=False): - if aux_id is not None: - aux_id = np.asarray(aux_id) - aux_id = torch.from_numpy(aux_id) - if cuda: - return aux_id.cuda() - return aux_id - - -def embedding_to_torch(d_vector, cuda=False): - if d_vector is not None: - d_vector = np.asarray(d_vector) - d_vector = torch.from_numpy(d_vector).float() - d_vector = d_vector.squeeze().unsqueeze(0) - if cuda: - return d_vector.cuda() - return d_vector - - -def numpy_to_torch(np_array, dtype, cuda=False): - if np_array is None: - return None - tensor = torch.as_tensor(np_array, dtype=dtype) - if cuda: - return tensor.cuda() - return tensor - - -def get_mask_from_lengths(lengths: torch.Tensor) -> torch.Tensor: - batch_size = lengths.shape[0] - max_len = torch.max(lengths).item() - ids = torch.arange(0, max_len, device=lengths.device).unsqueeze(0).expand(batch_size, -1) - mask = ids >= lengths.unsqueeze(1).expand(-1, max_len) - return mask - - -def pad(input_ele: List[torch.Tensor], max_len: int) -> torch.Tensor: - out_list = torch.jit.annotate(List[torch.Tensor], []) - for batch in input_ele: - if len(batch.shape) == 1: - one_batch_padded = F.pad(batch, (0, max_len - batch.size(0)), "constant", 0.0) - else: - one_batch_padded = F.pad(batch, (0, 0, 0, max_len - batch.size(0)), "constant", 0.0) - out_list.append(one_batch_padded) - out_padded = torch.stack(out_list) - return out_padded - - -def init_weights(m: nn.Module, mean: float = 0.0, std: float = 0.01): - classname = m.__class__.__name__ - if classname.find("Conv") != -1: - m.weight.data.normal_(mean, std) - - -def stride_lens(lens: torch.Tensor, stride: int = 2) -> torch.Tensor: - return torch.ceil(lens / stride).int() - - -def initialize_embeddings(shape: Tuple[int]) -> torch.Tensor: - assert len(shape) == 2, "Can only initialize 2-D embedding matrices ..." - return torch.randn(shape) * np.sqrt(2 / shape[1]) - - -# pylint: disable=redefined-outer-name -def calc_same_padding(kernel_size: int) -> Tuple[int, int]: - pad = kernel_size // 2 - return (pad, pad - (kernel_size + 1) % 2) +logger = logging.getLogger(__name__) hann_window = {} mel_basis = {} -@torch.no_grad() -def weights_reset(m: nn.Module): - # check if the current module has reset_parameters and if it is reset the weight - reset_parameters = getattr(m, "reset_parameters", None) - if callable(reset_parameters): - m.reset_parameters() - - -def get_module_weights_sum(mdl: nn.Module): - dict_sums = {} - for name, w in mdl.named_parameters(): - if "weight" in name: - value = w.data.sum().item() - dict_sums[name] = value - return dict_sums - - -def load_audio(file_path: str): - """Load the audio file normalized in [-1, 1] - - Return Shapes: - - x: :math:`[1, T]` - """ - x, sr = torchaudio.load( - file_path, - ) - assert (x > 1).sum() + (x < -1).sum() == 0 - return x, sr - - -def _amp_to_db(x, C=1, clip_val=1e-5): - return torch.log(torch.clamp(x, min=clip_val) * C) - - -def _db_to_amp(x, C=1): - return torch.exp(x) / C - - -def amp_to_db(magnitudes): - output = _amp_to_db(magnitudes) - return output - - -def db_to_amp(magnitudes): - output = _db_to_amp(magnitudes) - return output - - -def _wav_to_spec(y, n_fft, hop_length, win_length, center=False): - y = y.squeeze(1) - - if torch.min(y) < -1.0: - print("min value is ", torch.min(y)) - if torch.max(y) > 1.0: - print("max value is ", torch.max(y)) - - global hann_window # pylint: disable=global-statement - dtype_device = str(y.dtype) + "_" + str(y.device) - wnsize_dtype_device = str(win_length) + "_" + dtype_device - if wnsize_dtype_device not in hann_window: - hann_window[wnsize_dtype_device] = torch.hann_window(win_length).to(dtype=y.dtype, device=y.device) - - y = torch.nn.functional.pad( - y.unsqueeze(1), - (int((n_fft - hop_length) / 2), int((n_fft - hop_length) / 2)), - mode="reflect", - ) - y = y.squeeze(1) - - spec = torch.stft( - y, - n_fft, - hop_length=hop_length, - win_length=win_length, - window=hann_window[wnsize_dtype_device], - center=center, - pad_mode="reflect", - normalized=False, - onesided=True, - return_complex=False, - ) - - return spec - - -def wav_to_spec(y, n_fft, hop_length, win_length, center=False): - """ - Args Shapes: - - y : :math:`[B, 1, T]` - - Return Shapes: - - spec : :math:`[B,C,T]` - """ - spec = _wav_to_spec(y, n_fft, hop_length, win_length, center=center) - spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) - return spec - - def wav_to_energy(y, n_fft, hop_length, win_length, center=False): - spec = _wav_to_spec(y, n_fft, hop_length, win_length, center=center) - - spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) + spec = wav_to_spec(y, n_fft, hop_length, win_length, center=center) return torch.norm(spec, dim=1, keepdim=True) -def name_mel_basis(spec, n_fft, fmax): - n_fft_len = f"{n_fft}_{fmax}_{spec.dtype}_{spec.device}" - return n_fft_len - - -def spec_to_mel(spec, n_fft, num_mels, sample_rate, fmin, fmax): - """ - Args Shapes: - - spec : :math:`[B,C,T]` - - Return Shapes: - - mel : :math:`[B,C,T]` - """ - global mel_basis # pylint: disable=global-statement - mel_basis_key = name_mel_basis(spec, n_fft, fmax) - # pylint: disable=too-many-function-args - if mel_basis_key not in mel_basis: - # pylint: disable=missing-kwoa - mel = librosa_mel_fn(sample_rate, n_fft, num_mels, fmin, fmax) - mel_basis[mel_basis_key] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device) - mel = torch.matmul(mel_basis[mel_basis_key], spec) - mel = amp_to_db(mel) - return mel - - -def wav_to_mel(y, n_fft, num_mels, sample_rate, hop_length, win_length, fmin, fmax, center=False): - """ - Args Shapes: - - y : :math:`[B, 1, T_y]` - - Return Shapes: - - spec : :math:`[B,C,T_spec]` - """ - y = y.squeeze(1) - - if torch.min(y) < -1.0: - print("min value is ", torch.min(y)) - if torch.max(y) > 1.0: - print("max value is ", torch.max(y)) - - global mel_basis, hann_window # pylint: disable=global-statement - mel_basis_key = name_mel_basis(y, n_fft, fmax) - wnsize_dtype_device = str(win_length) + "_" + str(y.dtype) + "_" + str(y.device) - if mel_basis_key not in mel_basis: - # pylint: disable=missing-kwoa - mel = librosa_mel_fn( - sr=sample_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax - ) # pylint: disable=too-many-function-args - mel_basis[mel_basis_key] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device) - if wnsize_dtype_device not in hann_window: - hann_window[wnsize_dtype_device] = torch.hann_window(win_length).to(dtype=y.dtype, device=y.device) - - y = torch.nn.functional.pad( - y.unsqueeze(1), - (int((n_fft - hop_length) / 2), int((n_fft - hop_length) / 2)), - mode="reflect", - ) - y = y.squeeze(1) - - spec = torch.stft( - y, - n_fft, - hop_length=hop_length, - win_length=win_length, - window=hann_window[wnsize_dtype_device], - center=center, - pad_mode="reflect", - normalized=False, - onesided=True, - return_complex=False, - ) - - spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) - spec = torch.matmul(mel_basis[mel_basis_key], spec) - spec = amp_to_db(spec) - return spec - - ############################## # DATASET ############################## -def get_attribute_balancer_weights(items: list, attr_name: str, multi_dict: dict = None): - """Create balancer weight for torch WeightedSampler""" - attr_names_samples = np.array([item[attr_name] for item in items]) - unique_attr_names = np.unique(attr_names_samples).tolist() - attr_idx = [unique_attr_names.index(l) for l in attr_names_samples] - attr_count = np.array([len(np.where(attr_names_samples == l)[0]) for l in unique_attr_names]) - weight_attr = 1.0 / attr_count - dataset_samples_weight = np.array([weight_attr[l] for l in attr_idx]) - dataset_samples_weight = dataset_samples_weight / np.linalg.norm(dataset_samples_weight) - if multi_dict is not None: - multiplier_samples = np.array([multi_dict.get(item[attr_name], 1.0) for item in items]) - dataset_samples_weight *= multiplier_samples - return ( - torch.from_numpy(dataset_samples_weight).float(), - unique_attr_names, - np.unique(dataset_samples_weight).tolist(), - ) - - class ForwardTTSE2eF0Dataset(F0Dataset): """Override F0Dataset to avoid slow computing of pitches""" @@ -324,7 +66,6 @@ def __init__( self, ap, samples: Union[List[List], List[Dict]], - verbose=False, cache_path: str = None, precompute_num_workers=0, normalize_f0=True, @@ -332,7 +73,6 @@ def __init__( super().__init__( samples=samples, ap=ap, - verbose=verbose, cache_path=cache_path, precompute_num_workers=precompute_num_workers, normalize_f0=normalize_f0, @@ -404,7 +144,7 @@ def __getitem__(self, idx): try: token_ids = self.get_token_ids(idx, item["text"]) except: - print(idx, item) + logger.exception("%s %s", idx, item) # pylint: disable=raise-missing-from raise OSError f0 = None @@ -769,7 +509,7 @@ def init_multispeaker(self, config: Coqpit): def _init_speaker_embedding(self): # pylint: disable=attribute-defined-outside-init if self.num_speakers > 0: - print(" > initialization of speaker-embedding layers.") + logger.info("Initialization of speaker-embedding layers.") self.embedded_speaker_dim = self.args.speaker_embedding_channels self.args.embedded_speaker_dim = self.args.speaker_embedding_channels @@ -953,7 +693,7 @@ def train_step(self, batch: dict, criterion: nn.Module, optimizer_idx: int): ) # compute loss - with autocast(enabled=False): # use float32 for the criterion + with torch.autocast("cuda", enabled=False): # use float32 for the criterion loss_dict = criterion[optimizer_idx]( scores_disc_fake=scores_d_fake, scores_disc_real=scores_d_real, @@ -964,7 +704,7 @@ def train_step(self, batch: dict, criterion: nn.Module, optimizer_idx: int): if optimizer_idx == 1: mel = batch["mel_input"] # compute melspec segment - with autocast(enabled=False): + with torch.autocast("cuda", enabled=False): mel_slice = segment( mel.float(), self.model_outputs_cache["slice_ids"], self.args.spec_segment_size, pad_short=True ) @@ -992,7 +732,7 @@ def train_step(self, batch: dict, criterion: nn.Module, optimizer_idx: int): ) # compute losses - with autocast(enabled=True): # use float32 for the criterion + with torch.autocast("cuda", enabled=True): # use float32 for the criterion loss_dict = criterion[optimizer_idx]( mel_output=self.model_outputs_cache["acoustic_model_outputs"].transpose(1, 2), mel_target=batch["mel_input"], @@ -1198,7 +938,7 @@ def synthesize( **kwargs, ): # pylint: disable=unused-argument # TODO: add cloning support with ref_waveform - is_cuda = next(self.parameters()).is_cuda + device = next(self.parameters()).device # convert text to sequence of token IDs text_inputs = np.asarray( @@ -1212,14 +952,14 @@ def synthesize( if isinstance(speaker_id, str) and self.args.use_speaker_embedding: # get the speaker id for the speaker embedding layer _speaker_id = self.speaker_manager.name_to_id[speaker_id] - _speaker_id = id_to_torch(_speaker_id, cuda=is_cuda) + _speaker_id = id_to_torch(_speaker_id, device=device) if speaker_id is not None and self.args.use_d_vector_file: # get the average d_vector for the speaker d_vector = self.speaker_manager.get_mean_embedding(speaker_id, num_samples=None, randomize=False) - d_vector = embedding_to_torch(d_vector, cuda=is_cuda) + d_vector = embedding_to_torch(d_vector, device=device) - text_inputs = numpy_to_torch(text_inputs, torch.long, cuda=is_cuda) + text_inputs = numpy_to_torch(text_inputs, torch.long, device=device) text_inputs = text_inputs.unsqueeze(0) # synthesize voice @@ -1242,7 +982,7 @@ def synthesize( return return_dict def synthesize_with_gl(self, text: str, speaker_id, d_vector): - is_cuda = next(self.parameters()).is_cuda + device = next(self.parameters()).device # convert text to sequence of token IDs text_inputs = np.asarray( @@ -1251,12 +991,12 @@ def synthesize_with_gl(self, text: str, speaker_id, d_vector): ) # pass tensors to backend if speaker_id is not None: - speaker_id = id_to_torch(speaker_id, cuda=is_cuda) + speaker_id = id_to_torch(speaker_id, device=device) if d_vector is not None: - d_vector = embedding_to_torch(d_vector, cuda=is_cuda) + d_vector = embedding_to_torch(d_vector, device=device) - text_inputs = numpy_to_torch(text_inputs, torch.long, cuda=is_cuda) + text_inputs = numpy_to_torch(text_inputs, torch.long, device=device) text_inputs = text_inputs.unsqueeze(0) # synthesize voice @@ -1287,7 +1027,7 @@ def test_run(self, assets) -> Tuple[Dict, Dict]: Returns: Tuple[Dict, Dict]: Test figures and audios to be projected to Tensorboard. """ - print(" | > Synthesizing test sentences.") + logger.info("Synthesizing test sentences.") test_audios = {} test_figures = {} test_sentences = self.config.test_sentences @@ -1401,14 +1141,14 @@ def get_sampler(self, config: Coqpit, dataset: TTSDataset, num_gpus=1): data_items = dataset.samples if getattr(config, "use_weighted_sampler", False): for attr_name, alpha in config.weighted_sampler_attrs.items(): - print(f" > Using weighted sampler for attribute '{attr_name}' with alpha '{alpha}'") + logger.info("Using weighted sampler for attribute '%s' with alpha %.2f", attr_name, alpha) multi_dict = config.weighted_sampler_multipliers.get(attr_name, None) - print(multi_dict) + logger.info(multi_dict) weights, attr_names, attr_weights = get_attribute_balancer_weights( attr_name=attr_name, items=data_items, multi_dict=multi_dict ) weights = weights * alpha - print(f" > Attribute weights for '{attr_names}' \n | > {attr_weights}") + logger.info("Attribute weights for '%s' \n | > %s", attr_names, attr_weights) if weights is not None: sampler = WeightedRandomSampler(weights, len(weights)) @@ -1448,7 +1188,6 @@ def get_data_loader( compute_f0=config.compute_f0, f0_cache_path=config.f0_cache_path, attn_prior_cache_path=config.attn_prior_cache_path if config.use_attn_priors else None, - verbose=verbose, tokenizer=self.tokenizer, start_by_longest=config.start_by_longest, ) @@ -1525,7 +1264,7 @@ def on_epoch_end(self, trainer): # pylint: disable=unused-argument @staticmethod def init_from_config( - config: "DelightfulTTSConfig", samples: Union[List[List], List[Dict]] = None, verbose=False + config: "DelightfulTTSConfig", samples: Union[List[List], List[Dict]] = None ): # pylint: disable=unused-argument """Initiate model from config @@ -1604,36 +1343,6 @@ def __init__(self, config): self.gen_loss_alpha = config.gen_loss_alpha self.multi_scale_stft_loss_alpha = config.multi_scale_stft_loss_alpha - @staticmethod - def _binary_alignment_loss(alignment_hard, alignment_soft): - """Binary loss that forces soft alignments to match the hard alignments as - explained in `https://arxiv.org/pdf/2108.10447.pdf`. - """ - log_sum = torch.log(torch.clamp(alignment_soft[alignment_hard == 1], min=1e-12)).sum() - return -log_sum / alignment_hard.sum() - - @staticmethod - def feature_loss(feats_real, feats_generated): - loss = 0 - for dr, dg in zip(feats_real, feats_generated): - for rl, gl in zip(dr, dg): - rl = rl.float().detach() - gl = gl.float() - loss += torch.mean(torch.abs(rl - gl)) - return loss * 2 - - @staticmethod - def generator_loss(scores_fake): - loss = 0 - gen_losses = [] - for dg in scores_fake: - dg = dg.float() - l = torch.mean((1 - dg) ** 2) - gen_losses.append(l) - loss += l - - return loss, gen_losses - def forward( self, mel_output, @@ -1731,7 +1440,7 @@ def forward( ) if self.binary_alignment_loss_alpha > 0 and aligner_hard is not None: - binary_alignment_loss = self._binary_alignment_loss(aligner_hard, aligner_soft) + binary_alignment_loss = _binary_alignment_loss(aligner_hard, aligner_soft) total_loss = total_loss + self.binary_alignment_loss_alpha * binary_alignment_loss * binary_loss_weight if binary_loss_weight: loss_dict["loss_binary_alignment"] = ( @@ -1751,8 +1460,8 @@ def forward( # vocoder losses if not skip_disc: - loss_feat = self.feature_loss(feats_real=feats_real, feats_generated=feats_fake) * self.feat_loss_alpha - loss_gen = self.generator_loss(scores_fake=scores_fake)[0] * self.gen_loss_alpha + loss_feat = feature_loss(feats_real=feats_real, feats_generated=feats_fake) * self.feat_loss_alpha + loss_gen = generator_loss(scores_fake=scores_fake)[0] * self.gen_loss_alpha loss_dict["vocoder_loss_feat"] = loss_feat loss_dict["vocoder_loss_gen"] = loss_gen loss_dict["loss"] = loss_dict["loss"] + loss_feat + loss_gen diff --git a/TTS/tts/models/forward_tts.py b/TTS/tts/models/forward_tts.py index b6e9ac8a14..d09e3ea91b 100644 --- a/TTS/tts/models/forward_tts.py +++ b/TTS/tts/models/forward_tts.py @@ -1,10 +1,12 @@ +import logging from dataclasses import dataclass, field from typing import Dict, List, Tuple, Union import torch from coqpit import Coqpit +from monotonic_alignment_search import maximum_path from torch import nn -from torch.cuda.amp.autocast_mode import autocast +from trainer.io import load_fsspec from TTS.tts.layers.feed_forward.decoder import Decoder from TTS.tts.layers.feed_forward.encoder import Encoder @@ -12,11 +14,12 @@ from TTS.tts.layers.generic.pos_encoding import PositionalEncoding from TTS.tts.layers.glow_tts.duration_predictor import DurationPredictor from TTS.tts.models.base_tts import BaseTTS -from TTS.tts.utils.helpers import average_over_durations, generate_path, maximum_path, sequence_mask +from TTS.tts.utils.helpers import average_over_durations, expand_encoder_outputs, generate_attention, sequence_mask from TTS.tts.utils.speakers import SpeakerManager from TTS.tts.utils.text.tokenizer import TTSTokenizer from TTS.tts.utils.visual import plot_alignment, plot_avg_energy, plot_avg_pitch, plot_spectrogram -from TTS.utils.io import load_fsspec + +logger = logging.getLogger(__name__) @dataclass @@ -299,57 +302,14 @@ def init_multispeaker(self, config: Coqpit): if config.use_d_vector_file: self.embedded_speaker_dim = config.d_vector_dim if self.args.d_vector_dim != self.args.hidden_channels: - #self.proj_g = nn.Conv1d(self.args.d_vector_dim, self.args.hidden_channels, 1) + # self.proj_g = nn.Conv1d(self.args.d_vector_dim, self.args.hidden_channels, 1) self.proj_g = nn.Linear(in_features=self.args.d_vector_dim, out_features=self.args.hidden_channels) # init speaker embedding layer if config.use_speaker_embedding and not config.use_d_vector_file: - print(" > Init speaker_embedding layer.") + logger.info("Init speaker_embedding layer.") self.emb_g = nn.Embedding(self.num_speakers, self.args.hidden_channels) nn.init.uniform_(self.emb_g.weight, -0.1, 0.1) - @staticmethod - def generate_attn(dr, x_mask, y_mask=None): - """Generate an attention mask from the durations. - - Shapes - - dr: :math:`(B, T_{en})` - - x_mask: :math:`(B, T_{en})` - - y_mask: :math:`(B, T_{de})` - """ - # compute decode mask from the durations - if y_mask is None: - y_lengths = dr.sum(1).long() - y_lengths[y_lengths < 1] = 1 - y_mask = torch.unsqueeze(sequence_mask(y_lengths, None), 1).to(dr.dtype) - attn_mask = torch.unsqueeze(x_mask, -1) * torch.unsqueeze(y_mask, 2) - attn = generate_path(dr, attn_mask.squeeze(1)).to(dr.dtype) - return attn - - def expand_encoder_outputs(self, en, dr, x_mask, y_mask): - """Generate attention alignment map from durations and - expand encoder outputs - - Shapes: - - en: :math:`(B, D_{en}, T_{en})` - - dr: :math:`(B, T_{en})` - - x_mask: :math:`(B, T_{en})` - - y_mask: :math:`(B, T_{de})` - - Examples:: - - encoder output: [a,b,c,d] - durations: [1, 3, 2, 1] - - expanded: [a, b, b, b, c, c, d] - attention map: [[0, 0, 0, 0, 0, 0, 1], - [0, 0, 0, 0, 1, 1, 0], - [0, 1, 1, 1, 0, 0, 0], - [1, 0, 0, 0, 0, 0, 0]] - """ - attn = self.generate_attn(dr, x_mask, y_mask) - o_en_ex = torch.matmul(attn.squeeze(1).transpose(1, 2).to(en.dtype), en.transpose(1, 2)).transpose(1, 2) - return o_en_ex, attn - def format_durations(self, o_dr_log, x_mask): """Format predicted durations. 1. Convert to linear scale from log scale @@ -404,13 +364,13 @@ def _forward_encoder( # [B, T, C] x_emb = self.emb(x) # encoder pass - #o_en = self.encoder(torch.transpose(x_emb, 1, -1), x_mask) + # o_en = self.encoder(torch.transpose(x_emb, 1, -1), x_mask) o_en = self.encoder(torch.transpose(x_emb, 1, -1), x_mask, g) # speaker conditioning # TODO: try different ways of conditioning - if g is not None: + if g is not None: if hasattr(self, "proj_g"): - g = self.proj_g(g.view(g.shape[0], -1)).unsqueeze(-1) + g = self.proj_g(g.view(g.shape[0], -1)).unsqueeze(-1) o_en = o_en + g return o_en, x_mask, g, x_emb @@ -440,9 +400,8 @@ def _forward_decoder( Returns: Tuple[torch.FloatTensor, torch.FloatTensor]: Decoder output, attention map from durations. """ - y_mask = torch.unsqueeze(sequence_mask(y_lengths, None), 1).to(o_en.dtype) # expand o_en with durations - o_en_ex, attn = self.expand_encoder_outputs(o_en, dr, x_mask, y_mask) + o_en_ex, attn, y_mask = expand_encoder_outputs(o_en, dr, x_mask, y_lengths) # positional encoding if hasattr(self, "pos_encoder"): o_en_ex = self.pos_encoder(o_en_ex, y_mask) @@ -621,7 +580,7 @@ def forward( o_dr_log = self.duration_predictor(o_en, x_mask) o_dr = torch.clamp(torch.exp(o_dr_log) - 1, 0, self.max_duration) # generate attn mask from predicted durations - o_attn = self.generate_attn(o_dr.squeeze(1), x_mask) + o_attn = generate_attention(o_dr.squeeze(1), x_mask) # aligner o_alignment_dur = None alignment_soft = None @@ -740,7 +699,7 @@ def train_step(self, batch: dict, criterion: nn.Module): if self.use_aligner: durations = outputs["o_alignment_dur"] # use float32 in AMP - with autocast(enabled=False): + with torch.autocast("cuda", enabled=False): # compute loss loss_dict = criterion( decoder_output=outputs["model_outputs"], diff --git a/TTS/tts/models/glow_tts.py b/TTS/tts/models/glow_tts.py index bfd1a2b618..5bf4713140 100644 --- a/TTS/tts/models/glow_tts.py +++ b/TTS/tts/models/glow_tts.py @@ -1,22 +1,25 @@ +import logging import math from typing import Dict, List, Tuple, Union import torch from coqpit import Coqpit +from monotonic_alignment_search import maximum_path from torch import nn -from torch.cuda.amp.autocast_mode import autocast from torch.nn import functional as F +from trainer.io import load_fsspec from TTS.tts.configs.glow_tts_config import GlowTTSConfig from TTS.tts.layers.glow_tts.decoder import Decoder from TTS.tts.layers.glow_tts.encoder import Encoder from TTS.tts.models.base_tts import BaseTTS -from TTS.tts.utils.helpers import generate_path, maximum_path, sequence_mask +from TTS.tts.utils.helpers import generate_path, sequence_mask from TTS.tts.utils.speakers import SpeakerManager from TTS.tts.utils.synthesis import synthesis from TTS.tts.utils.text.tokenizer import TTSTokenizer from TTS.tts.utils.visual import plot_alignment, plot_spectrogram -from TTS.utils.io import load_fsspec + +logger = logging.getLogger(__name__) class GlowTTS(BaseTTS): @@ -53,7 +56,7 @@ class GlowTTS(BaseTTS): >>> from TTS.tts.configs.glow_tts_config import GlowTTSConfig >>> from TTS.tts.models.glow_tts import GlowTTS >>> config = GlowTTSConfig() - >>> model = GlowTTS.init_from_config(config, verbose=False) + >>> model = GlowTTS.init_from_config(config) """ def __init__( @@ -127,7 +130,7 @@ def init_multispeaker(self, config: Coqpit): ), " [!] d-vector dimension mismatch b/w config and speaker manager." # init speaker embedding layer if config.use_speaker_embedding and not config.use_d_vector_file: - print(" > Init speaker_embedding layer.") + logger.info("Init speaker_embedding layer.") self.embedded_speaker_dim = self.hidden_channels_enc self.emb_g = nn.Embedding(self.num_speakers, self.hidden_channels_enc) nn.init.uniform_(self.emb_g.weight, -0.1, 0.1) @@ -412,7 +415,7 @@ def train_step(self, batch: dict, criterion: nn.Module): aux_input={"d_vectors": d_vectors, "speaker_ids": speaker_ids}, ) - with autocast(enabled=False): # avoid mixed_precision in criterion + with torch.autocast("cuda", enabled=False): # avoid mixed_precision in criterion loss_dict = criterion( outputs["z"].float(), outputs["y_mean"].float(), @@ -479,13 +482,13 @@ def test_run(self, assets: Dict) -> Tuple[Dict, Dict]: Returns: Tuple[Dict, Dict]: Test figures and audios to be projected to Tensorboard. """ - print(" | > Synthesizing test sentences.") + logger.info("Synthesizing test sentences.") test_audios = {} test_figures = {} test_sentences = self.config.test_sentences aux_inputs = self._get_test_aux_input() if len(test_sentences) == 0: - print(" | [!] No test sentences provided.") + logger.warning("No test sentences provided.") else: for idx, sen in enumerate(test_sentences): outputs = synthesis( @@ -540,18 +543,17 @@ def on_train_step_start(self, trainer): self.run_data_dep_init = trainer.total_steps_done < self.data_dep_init_steps @staticmethod - def init_from_config(config: "GlowTTSConfig", samples: Union[List[List], List[Dict]] = None, verbose=True): + def init_from_config(config: "GlowTTSConfig", samples: Union[List[List], List[Dict]] = None): """Initiate model from config Args: config (VitsConfig): Model config. samples (Union[List[List], List[Dict]]): Training samples to parse speaker ids for training. Defaults to None. - verbose (bool): If True, print init messages. Defaults to True. """ from TTS.utils.audio import AudioProcessor - ap = AudioProcessor.init_from_config(config, verbose) + ap = AudioProcessor.init_from_config(config) tokenizer, new_config = TTSTokenizer.init_from_config(config) speaker_manager = SpeakerManager.init_from_config(config, samples) return GlowTTS(new_config, ap, tokenizer, speaker_manager) diff --git a/TTS/tts/models/neuralhmm_tts.py b/TTS/tts/models/neuralhmm_tts.py index e241410872..0b3fadafbf 100644 --- a/TTS/tts/models/neuralhmm_tts.py +++ b/TTS/tts/models/neuralhmm_tts.py @@ -1,11 +1,14 @@ +import logging import os from typing import Dict, List, Union import torch from coqpit import Coqpit from torch import nn +from trainer.io import load_fsspec from trainer.logging.tensorboard_logger import TensorboardLogger +from TTS.tts.layers.losses import NLLLoss from TTS.tts.layers.overflow.common_layers import Encoder, OverflowUtils from TTS.tts.layers.overflow.neural_hmm import NeuralHMM from TTS.tts.layers.overflow.plotting_utils import ( @@ -16,8 +19,9 @@ from TTS.tts.utils.speakers import SpeakerManager from TTS.tts.utils.text.tokenizer import TTSTokenizer from TTS.tts.utils.visual import plot_alignment, plot_spectrogram -from TTS.utils.generic_utils import format_aux_input -from TTS.utils.io import load_fsspec +from TTS.utils.generic_utils import format_aux_input, is_pytorch_at_least_2_4 + +logger = logging.getLogger(__name__) class NeuralhmmTTS(BaseTTS): @@ -104,7 +108,7 @@ def update_mean_std(self, statistics_dict: Dict): def preprocess_batch(self, text, text_len, mels, mel_len): if self.mean.item() == 0 or self.std.item() == 1: - statistics_dict = torch.load(self.mel_statistics_parameter_path) + statistics_dict = torch.load(self.mel_statistics_parameter_path, weights_only=is_pytorch_at_least_2_4()) self.update_mean_std(statistics_dict) mels = self.normalize(mels) @@ -235,18 +239,17 @@ def get_criterion(): return NLLLoss() @staticmethod - def init_from_config(config: "NeuralhmmTTSConfig", samples: Union[List[List], List[Dict]] = None, verbose=True): + def init_from_config(config: "NeuralhmmTTSConfig", samples: Union[List[List], List[Dict]] = None): """Initiate model from config Args: config (VitsConfig): Model config. samples (Union[List[List], List[Dict]]): Training samples to parse speaker ids for training. Defaults to None. - verbose (bool): If True, print init messages. Defaults to True. """ from TTS.utils.audio import AudioProcessor - ap = AudioProcessor.init_from_config(config, verbose) + ap = AudioProcessor.init_from_config(config) tokenizer, new_config = TTSTokenizer.init_from_config(config) speaker_manager = SpeakerManager.init_from_config(config, samples) return NeuralhmmTTS(new_config, ap, tokenizer, speaker_manager) @@ -266,14 +269,17 @@ def on_init_start(self, trainer): dataloader = trainer.get_train_dataloader( training_assets=None, samples=trainer.train_samples, verbose=False ) - print( - f" | > Data parameters not found for: {trainer.config.mel_statistics_parameter_path}. Computing mel normalization parameters..." + logger.info( + "Data parameters not found for: %s. Computing mel normalization parameters...", + trainer.config.mel_statistics_parameter_path, ) data_mean, data_std, init_transition_prob = OverflowUtils.get_data_parameters_for_flat_start( dataloader, trainer.config.out_channels, trainer.config.state_per_phone ) - print( - f" | > Saving data parameters to: {trainer.config.mel_statistics_parameter_path}: value: {data_mean, data_std, init_transition_prob}" + logger.info( + "Saving data parameters to: %s: value: %s", + trainer.config.mel_statistics_parameter_path, + (data_mean, data_std, init_transition_prob), ) statistics = { "mean": data_mean.item(), @@ -283,16 +289,19 @@ def on_init_start(self, trainer): torch.save(statistics, trainer.config.mel_statistics_parameter_path) else: - print( - f" | > Data parameters found for: {trainer.config.mel_statistics_parameter_path}. Loading mel normalization parameters..." + logger.info( + "Data parameters found for: %s. Loading mel normalization parameters...", + trainer.config.mel_statistics_parameter_path, + ) + statistics = torch.load( + trainer.config.mel_statistics_parameter_path, weights_only=is_pytorch_at_least_2_4() ) - statistics = torch.load(trainer.config.mel_statistics_parameter_path) data_mean, data_std, init_transition_prob = ( statistics["mean"], statistics["std"], statistics["init_transition_prob"], ) - print(f" | > Data parameters loaded with value: {data_mean, data_std, init_transition_prob}") + logger.info("Data parameters loaded with value: %s", (data_mean, data_std, init_transition_prob)) trainer.config.flat_start_params["transition_p"] = ( init_transition_prob.item() if torch.is_tensor(init_transition_prob) else init_transition_prob @@ -318,7 +327,7 @@ def _create_logs(self, batch, outputs, ap): # pylint: disable=no-self-use, unus } # sample one item from the batch -1 will give the smalles item - print(" | > Synthesising audio from the model...") + logger.info("Synthesising audio from the model...") inference_output = self.inference( batch["text_input"][-1].unsqueeze(0), aux_input={"x_lengths": batch["text_lengths"][-1].unsqueeze(0)} ) @@ -365,21 +374,3 @@ def test_log( ) -> None: logger.test_audios(steps, outputs[1], self.ap.sample_rate) logger.test_figures(steps, outputs[0]) - - -class NLLLoss(nn.Module): - """Negative log likelihood loss.""" - - def forward(self, log_prob: torch.Tensor) -> dict: # pylint: disable=no-self-use - """Compute the loss. - - Args: - logits (Tensor): [B, T, D] - - Returns: - Tensor: [1] - - """ - return_dict = {} - return_dict["loss"] = -log_prob.mean() - return return_dict diff --git a/TTS/tts/models/overflow.py b/TTS/tts/models/overflow.py index 92b3c767de..1c146b2eac 100644 --- a/TTS/tts/models/overflow.py +++ b/TTS/tts/models/overflow.py @@ -1,11 +1,14 @@ +import logging import os from typing import Dict, List, Union import torch from coqpit import Coqpit from torch import nn +from trainer.io import load_fsspec from trainer.logging.tensorboard_logger import TensorboardLogger +from TTS.tts.layers.losses import NLLLoss from TTS.tts.layers.overflow.common_layers import Encoder, OverflowUtils from TTS.tts.layers.overflow.decoder import Decoder from TTS.tts.layers.overflow.neural_hmm import NeuralHMM @@ -17,8 +20,9 @@ from TTS.tts.utils.speakers import SpeakerManager from TTS.tts.utils.text.tokenizer import TTSTokenizer from TTS.tts.utils.visual import plot_alignment, plot_spectrogram -from TTS.utils.generic_utils import format_aux_input -from TTS.utils.io import load_fsspec +from TTS.utils.generic_utils import format_aux_input, is_pytorch_at_least_2_4 + +logger = logging.getLogger(__name__) class Overflow(BaseTTS): @@ -29,32 +33,33 @@ class Overflow(BaseTTS): Paper abstract:: Neural HMMs are a type of neural transducer recently proposed for - sequence-to-sequence modelling in text-to-speech. They combine the best features - of classic statistical speech synthesis and modern neural TTS, requiring less - data and fewer training updates, and are less prone to gibberish output caused - by neural attention failures. In this paper, we combine neural HMM TTS with - normalising flows for describing the highly non-Gaussian distribution of speech - acoustics. The result is a powerful, fully probabilistic model of durations and - acoustics that can be trained using exact maximum likelihood. Compared to - dominant flow-based acoustic models, our approach integrates autoregression for - improved modelling of long-range dependences such as utterance-level prosody. - Experiments show that a system based on our proposal gives more accurate - pronunciations and better subjective speech quality than comparable methods, - whilst retaining the original advantages of neural HMMs. Audio examples and code - are available at https://shivammehta25.github.io/OverFlow/. + sequence-to-sequence modelling in text-to-speech. They combine the best features + of classic statistical speech synthesis and modern neural TTS, requiring less + data and fewer training updates, and are less prone to gibberish output caused + by neural attention failures. In this paper, we combine neural HMM TTS with + normalising flows for describing the highly non-Gaussian distribution of speech + acoustics. The result is a powerful, fully probabilistic model of durations and + acoustics that can be trained using exact maximum likelihood. Compared to + dominant flow-based acoustic models, our approach integrates autoregression for + improved modelling of long-range dependences such as utterance-level prosody. + Experiments show that a system based on our proposal gives more accurate + pronunciations and better subjective speech quality than comparable methods, + whilst retaining the original advantages of neural HMMs. Audio examples and code + are available at https://shivammehta25.github.io/OverFlow/. Note: - - Neural HMMs uses flat start initialization i.e it computes the means and std and transition probabilities - of the dataset and uses them to initialize the model. This benefits the model and helps with faster learning - If you change the dataset or want to regenerate the parameters change the `force_generate_statistics` and - `mel_statistics_parameter_path` accordingly. + - Neural HMMs uses flat start initialization i.e it computes the means + and std and transition probabilities of the dataset and uses them to initialize + the model. This benefits the model and helps with faster learning If you change + the dataset or want to regenerate the parameters change the + `force_generate_statistics` and `mel_statistics_parameter_path` accordingly. - To enable multi-GPU training, set the `use_grad_checkpointing=False` in config. - This will significantly increase the memory usage. This is because to compute - the actual data likelihood (not an approximation using MAS/Viterbi) we must use - all the states at the previous time step during the forward pass to decide the - probability distribution at the current step i.e the difference between the forward - algorithm and viterbi approximation. + This will significantly increase the memory usage. This is because to compute + the actual data likelihood (not an approximation using MAS/Viterbi) we must use + all the states at the previous time step during the forward pass to decide the + probability distribution at the current step i.e the difference between the forward + algorithm and viterbi approximation. Check :class:`TTS.tts.configs.overflow.OverFlowConfig` for class arguments. """ @@ -117,7 +122,7 @@ def update_mean_std(self, statistics_dict: Dict): def preprocess_batch(self, text, text_len, mels, mel_len): if self.mean.item() == 0 or self.std.item() == 1: - statistics_dict = torch.load(self.mel_statistics_parameter_path) + statistics_dict = torch.load(self.mel_statistics_parameter_path, weights_only=is_pytorch_at_least_2_4()) self.update_mean_std(statistics_dict) mels = self.normalize(mels) @@ -250,18 +255,17 @@ def get_criterion(): return NLLLoss() @staticmethod - def init_from_config(config: "OverFlowConfig", samples: Union[List[List], List[Dict]] = None, verbose=True): + def init_from_config(config: "OverFlowConfig", samples: Union[List[List], List[Dict]] = None): """Initiate model from config Args: config (VitsConfig): Model config. samples (Union[List[List], List[Dict]]): Training samples to parse speaker ids for training. Defaults to None. - verbose (bool): If True, print init messages. Defaults to True. """ from TTS.utils.audio import AudioProcessor - ap = AudioProcessor.init_from_config(config, verbose) + ap = AudioProcessor.init_from_config(config) tokenizer, new_config = TTSTokenizer.init_from_config(config) speaker_manager = SpeakerManager.init_from_config(config, samples) return Overflow(new_config, ap, tokenizer, speaker_manager) @@ -282,14 +286,17 @@ def on_init_start(self, trainer): dataloader = trainer.get_train_dataloader( training_assets=None, samples=trainer.train_samples, verbose=False ) - print( - f" | > Data parameters not found for: {trainer.config.mel_statistics_parameter_path}. Computing mel normalization parameters..." + logger.info( + "Data parameters not found for: %s. Computing mel normalization parameters...", + trainer.config.mel_statistics_parameter_path, ) data_mean, data_std, init_transition_prob = OverflowUtils.get_data_parameters_for_flat_start( dataloader, trainer.config.out_channels, trainer.config.state_per_phone ) - print( - f" | > Saving data parameters to: {trainer.config.mel_statistics_parameter_path}: value: {data_mean, data_std, init_transition_prob}" + logger.info( + "Saving data parameters to: %s: value: %s", + trainer.config.mel_statistics_parameter_path, + (data_mean, data_std, init_transition_prob), ) statistics = { "mean": data_mean.item(), @@ -299,16 +306,19 @@ def on_init_start(self, trainer): torch.save(statistics, trainer.config.mel_statistics_parameter_path) else: - print( - f" | > Data parameters found for: {trainer.config.mel_statistics_parameter_path}. Loading mel normalization parameters..." + logger.info( + "Data parameters found for: %s. Loading mel normalization parameters...", + trainer.config.mel_statistics_parameter_path, + ) + statistics = torch.load( + trainer.config.mel_statistics_parameter_path, weights_only=is_pytorch_at_least_2_4() ) - statistics = torch.load(trainer.config.mel_statistics_parameter_path) data_mean, data_std, init_transition_prob = ( statistics["mean"], statistics["std"], statistics["init_transition_prob"], ) - print(f" | > Data parameters loaded with value: {data_mean, data_std, init_transition_prob}") + logger.info("Data parameters loaded with value: %s", (data_mean, data_std, init_transition_prob)) trainer.config.flat_start_params["transition_p"] = ( init_transition_prob.item() if torch.is_tensor(init_transition_prob) else init_transition_prob @@ -334,7 +344,7 @@ def _create_logs(self, batch, outputs, ap): # pylint: disable=no-self-use, unus } # sample one item from the batch -1 will give the smalles item - print(" | > Synthesising audio from the model...") + logger.info("Synthesising audio from the model...") inference_output = self.inference( batch["text_input"][-1].unsqueeze(0), aux_input={"x_lengths": batch["text_lengths"][-1].unsqueeze(0)} ) @@ -381,21 +391,3 @@ def test_log( ) -> None: logger.test_audios(steps, outputs[1], self.ap.sample_rate) logger.test_figures(steps, outputs[0]) - - -class NLLLoss(nn.Module): - """Negative log likelihood loss.""" - - def forward(self, log_prob: torch.Tensor) -> dict: # pylint: disable=no-self-use - """Compute the loss. - - Args: - logits (Tensor): [B, T, D] - - Returns: - Tensor: [1] - - """ - return_dict = {} - return_dict["loss"] = -log_prob.mean() - return return_dict diff --git a/TTS/tts/models/tacotron.py b/TTS/tts/models/tacotron.py index 474ec4641d..5d3efd2021 100644 --- a/TTS/tts/models/tacotron.py +++ b/TTS/tts/models/tacotron.py @@ -4,7 +4,6 @@ import torch from torch import nn -from torch.cuda.amp.autocast_mode import autocast from trainer.trainer_utils import get_optimizer, get_scheduler from TTS.tts.layers.tacotron.capacitron_layers import CapacitronVAE @@ -101,12 +100,16 @@ def __init__( num_mel=self.decoder_output_dim, encoder_output_dim=self.encoder_in_features, capacitron_VAE_embedding_dim=self.capacitron_vae.capacitron_VAE_embedding_dim, - speaker_embedding_dim=self.embedded_speaker_dim - if self.use_speaker_embedding and self.capacitron_vae.capacitron_use_speaker_embedding - else None, - text_summary_embedding_dim=self.capacitron_vae.capacitron_text_summary_embedding_dim - if self.capacitron_vae.capacitron_use_text_summary_embeddings - else None, + speaker_embedding_dim=( + self.embedded_speaker_dim + if self.use_speaker_embedding and self.capacitron_vae.capacitron_use_speaker_embedding + else None + ), + text_summary_embedding_dim=( + self.capacitron_vae.capacitron_text_summary_embedding_dim + if self.capacitron_vae.capacitron_use_text_summary_embeddings + else None + ), ) # backward pass decoder @@ -171,9 +174,9 @@ def forward( # pylint: disable=dangerous-default-value encoder_outputs, *capacitron_vae_outputs = self.compute_capacitron_VAE_embedding( encoder_outputs, reference_mel_info=[mel_specs, mel_lengths], - text_info=[inputs, text_lengths] - if self.capacitron_vae.capacitron_use_text_summary_embeddings - else None, + text_info=( + [inputs, text_lengths] if self.capacitron_vae.capacitron_use_text_summary_embeddings else None + ), speaker_embedding=embedded_speakers if self.capacitron_vae.capacitron_use_speaker_embedding else None, ) else: @@ -237,13 +240,13 @@ def inference(self, text_input, aux_input=None): # B x capacitron_VAE_embedding_dim encoder_outputs, *_ = self.compute_capacitron_VAE_embedding( encoder_outputs, - reference_mel_info=[aux_input["style_mel"], reference_mel_length] - if aux_input["style_mel"] is not None - else None, + reference_mel_info=( + [aux_input["style_mel"], reference_mel_length] if aux_input["style_mel"] is not None else None + ), text_info=[style_text_embedding, style_text_length] if aux_input["style_text"] is not None else None, - speaker_embedding=aux_input["d_vectors"] - if self.capacitron_vae.capacitron_use_speaker_embedding - else None, + speaker_embedding=( + aux_input["d_vectors"] if self.capacitron_vae.capacitron_use_speaker_embedding else None + ), ) if self.num_speakers > 1: if not self.use_d_vector_file: @@ -306,7 +309,7 @@ def train_step(self, batch: Dict, criterion: torch.nn.Module) -> Tuple[Dict, Dic alignment_lengths = mel_lengths // self.decoder.r # compute loss - with autocast(enabled=False): # use float32 for the criterion + with torch.autocast("cuda", enabled=False): # use float32 for the criterion loss_dict = criterion( outputs["model_outputs"].float(), outputs["decoder_outputs"].float(), diff --git a/TTS/tts/models/tacotron2.py b/TTS/tts/models/tacotron2.py index 71ab1eac37..2716a39786 100644 --- a/TTS/tts/models/tacotron2.py +++ b/TTS/tts/models/tacotron2.py @@ -4,7 +4,6 @@ import torch from torch import nn -from torch.cuda.amp.autocast_mode import autocast from trainer.trainer_utils import get_optimizer, get_scheduler from TTS.tts.layers.tacotron.capacitron_layers import CapacitronVAE @@ -113,12 +112,14 @@ def __init__( num_mel=self.decoder_output_dim, encoder_output_dim=self.encoder_in_features, capacitron_VAE_embedding_dim=self.capacitron_vae.capacitron_VAE_embedding_dim, - speaker_embedding_dim=self.embedded_speaker_dim - if self.capacitron_vae.capacitron_use_speaker_embedding - else None, - text_summary_embedding_dim=self.capacitron_vae.capacitron_text_summary_embedding_dim - if self.capacitron_vae.capacitron_use_text_summary_embeddings - else None, + speaker_embedding_dim=( + self.embedded_speaker_dim if self.capacitron_vae.capacitron_use_speaker_embedding else None + ), + text_summary_embedding_dim=( + self.capacitron_vae.capacitron_text_summary_embedding_dim + if self.capacitron_vae.capacitron_use_text_summary_embeddings + else None + ), ) # backward pass decoder @@ -191,9 +192,11 @@ def forward( # pylint: disable=dangerous-default-value encoder_outputs, *capacitron_vae_outputs = self.compute_capacitron_VAE_embedding( encoder_outputs, reference_mel_info=[mel_specs, mel_lengths], - text_info=[embedded_inputs.transpose(1, 2), text_lengths] - if self.capacitron_vae.capacitron_use_text_summary_embeddings - else None, + text_info=( + [embedded_inputs.transpose(1, 2), text_lengths] + if self.capacitron_vae.capacitron_use_text_summary_embeddings + else None + ), speaker_embedding=embedded_speakers if self.capacitron_vae.capacitron_use_speaker_embedding else None, ) else: @@ -265,13 +268,13 @@ def inference(self, text, aux_input=None): # B x capacitron_VAE_embedding_dim encoder_outputs, *_ = self.compute_capacitron_VAE_embedding( encoder_outputs, - reference_mel_info=[aux_input["style_mel"], reference_mel_length] - if aux_input["style_mel"] is not None - else None, + reference_mel_info=( + [aux_input["style_mel"], reference_mel_length] if aux_input["style_mel"] is not None else None + ), text_info=[style_text_embedding, style_text_length] if aux_input["style_text"] is not None else None, - speaker_embedding=aux_input["d_vectors"] - if self.capacitron_vae.capacitron_use_speaker_embedding - else None, + speaker_embedding=( + aux_input["d_vectors"] if self.capacitron_vae.capacitron_use_speaker_embedding else None + ), ) if self.num_speakers > 1: @@ -334,7 +337,7 @@ def train_step(self, batch: Dict, criterion: torch.nn.Module): alignment_lengths = mel_lengths // self.decoder.r # compute loss - with autocast(enabled=False): # use float32 for the criterion + with torch.autocast("cuda", enabled=False): # use float32 for the criterion loss_dict = criterion( outputs["model_outputs"].float(), outputs["decoder_outputs"].float(), diff --git a/TTS/tts/models/tortoise.py b/TTS/tts/models/tortoise.py index 16644ff95e..738e9dd9b3 100644 --- a/TTS/tts/models/tortoise.py +++ b/TTS/tts/models/tortoise.py @@ -1,3 +1,4 @@ +import logging import os import random from contextlib import contextmanager @@ -22,6 +23,9 @@ from TTS.tts.layers.tortoise.vocoder import VocConf, VocType from TTS.tts.layers.tortoise.wav2vec_alignment import Wav2VecAlignment from TTS.tts.models.base_tts import BaseTTS +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 + +logger = logging.getLogger(__name__) def pad_or_truncate(t, length): @@ -100,7 +104,7 @@ def fix_autoregressive_output(codes, stop_token, complain=True): stop_token_indices = (codes == stop_token).nonzero() if len(stop_token_indices) == 0: if complain: - print( + logger.warning( "No stop tokens found in one of the generated voice clips. This typically means the spoken audio is " "too long. In some cases, the output will still be good, though. Listen to it and if it is missing words, " "try breaking up your input text." @@ -167,7 +171,13 @@ def classify_audio_clip(clip, model_dir): kernel_size=5, distribute_zero_label=False, ) - classifier.load_state_dict(torch.load(os.path.join(model_dir, "classifier.pth"), map_location=torch.device("cpu"))) + classifier.load_state_dict( + torch.load( + os.path.join(model_dir, "classifier.pth"), + map_location=torch.device("cpu"), + weights_only=is_pytorch_at_least_2_4(), + ) + ) clip = clip.cpu().unsqueeze(0) results = F.softmax(classifier(clip), dim=-1) return results[0][0] @@ -413,7 +423,9 @@ def get_conditioning_latents( Transforms one or more voice_samples into a tuple (autoregressive_conditioning_latent, diffusion_conditioning_latent). These are expressive learned latents that encode aspects of the provided clips like voice, intonation, and acoustic properties. - :param voice_samples: List of arbitrary reference clips, which should be *pairs* of torch tensors containing arbitrary kHz waveform data. + + :param voice_samples: List of arbitrary reference clips, which should be *pairs* + of torch tensors containing arbitrary kHz waveform data. :param latent_averaging_mode: 0/1/2 for following modes: 0 - latents will be generated as in original tortoise, using ~4.27s from each voice sample, averaging latent across all samples 1 - latents will be generated using (almost) entire voice samples, averaged across all the ~4.27s chunks @@ -485,6 +497,7 @@ def get_random_conditioning_latents(self): torch.load( os.path.join(self.models_dir, "rlg_auto.pth"), map_location=torch.device("cpu"), + weights_only=is_pytorch_at_least_2_4(), ) ) self.rlg_diffusion = RandomLatentConverter(2048).eval() @@ -492,6 +505,7 @@ def get_random_conditioning_latents(self): torch.load( os.path.join(self.models_dir, "rlg_diffuser.pth"), map_location=torch.device("cpu"), + weights_only=is_pytorch_at_least_2_4(), ) ) with torch.no_grad(): @@ -659,7 +673,7 @@ def inference( As cond_free_k increases, the output becomes dominated by the conditioning-free signal. diffusion_temperature: (float) Controls the variance of the noise fed into the diffusion model. [0,1]. Values at 0 are the "mean" prediction of the diffusion network and will sound bland and smeared. - hf_generate_kwargs: (**kwargs) The huggingface Transformers generate API is used for the autoregressive transformer. + hf_generate_kwargs: (`**kwargs`) The huggingface Transformers generate API is used for the autoregressive transformer. Extra keyword args fed to this function get forwarded directly to that API. Documentation here: https://huggingface.co/docs/transformers/internal/generation_utils @@ -713,10 +727,10 @@ def inference( 83 # This is the token for coding silence, which is fixed in place with "fix_autoregressive_output" ) self.autoregressive = self.autoregressive.to(self.device) - if verbose: - print("Generating autoregressive samples..") - with self.temporary_cuda(self.autoregressive) as autoregressive, torch.autocast( - device_type="cuda", dtype=torch.float16, enabled=half + logger.info("Generating autoregressive samples..") + with ( + self.temporary_cuda(self.autoregressive) as autoregressive, + torch.autocast(device_type="cuda", dtype=torch.float16, enabled=half), ): for b in tqdm(range(num_batches), disable=not verbose): codes = autoregressive.inference_speech( @@ -737,8 +751,9 @@ def inference( self.autoregressive_batch_size = orig_batch_size # in the case of single_sample clip_results = [] - with self.temporary_cuda(self.clvp) as clvp, torch.autocast( - device_type="cuda", dtype=torch.float16, enabled=half + with ( + self.temporary_cuda(self.clvp) as clvp, + torch.autocast(device_type="cuda", dtype=torch.float16, enabled=half), ): for batch in tqdm(samples, disable=not verbose): for i in range(batch.shape[0]): @@ -773,8 +788,7 @@ def inference( ) del auto_conditioning - if verbose: - print("Transforming autoregressive outputs into audio..") + logger.info("Transforming autoregressive outputs into audio..") wav_candidates = [] for b in range(best_results.shape[0]): codes = best_results[b].unsqueeze(0) @@ -878,17 +892,17 @@ def load_checkpoint( if os.path.exists(ar_path): # remove keys from the checkpoint that are not in the model - checkpoint = torch.load(ar_path, map_location=torch.device("cpu")) + checkpoint = torch.load(ar_path, map_location=torch.device("cpu"), weights_only=is_pytorch_at_least_2_4()) # strict set False # due to removed `bias` and `masked_bias` changes in Transformers self.autoregressive.load_state_dict(checkpoint, strict=False) if os.path.exists(diff_path): - self.diffusion.load_state_dict(torch.load(diff_path), strict=strict) + self.diffusion.load_state_dict(torch.load(diff_path, weights_only=is_pytorch_at_least_2_4()), strict=strict) if os.path.exists(clvp_path): - self.clvp.load_state_dict(torch.load(clvp_path), strict=strict) + self.clvp.load_state_dict(torch.load(clvp_path, weights_only=is_pytorch_at_least_2_4()), strict=strict) if os.path.exists(vocoder_checkpoint_path): self.vocoder.load_state_dict( @@ -896,6 +910,7 @@ def load_checkpoint( torch.load( vocoder_checkpoint_path, map_location=torch.device("cpu"), + weights_only=is_pytorch_at_least_2_4(), ) ) ) diff --git a/TTS/tts/models/vits.py b/TTS/tts/models/vits.py index d9b1f59618..7ec2519236 100644 --- a/TTS/tts/models/vits.py +++ b/TTS/tts/models/vits.py @@ -1,3 +1,4 @@ +import logging import math import os from dataclasses import dataclass, field, replace @@ -9,43 +10,41 @@ import torch.distributed as dist import torchaudio from coqpit import Coqpit -from librosa.filters import mel as librosa_mel_fn +from monotonic_alignment_search import maximum_path from torch import nn -from torch.cuda.amp.autocast_mode import autocast from torch.nn import functional as F from torch.utils.data import DataLoader from torch.utils.data.sampler import WeightedRandomSampler +from trainer.io import load_fsspec from trainer.torch import DistributedSampler, DistributedSamplerWrapper from trainer.trainer_utils import get_optimizer, get_scheduler from TTS.tts.configs.shared_configs import CharactersConfig -from TTS.tts.datasets.dataset import TTSDataset, _parse_sample +from TTS.tts.datasets.dataset import TTSDataset, _parse_sample, get_attribute_balancer_weights from TTS.tts.layers.glow_tts.duration_predictor import DurationPredictor from TTS.tts.layers.vits.discriminator import VitsDiscriminator from TTS.tts.layers.vits.networks import PosteriorEncoder, ResidualCouplingBlocks, TextEncoder from TTS.tts.layers.vits.stochastic_duration_predictor import StochasticDurationPredictor from TTS.tts.models.base_tts import BaseTTS from TTS.tts.utils.fairseq import rehash_fairseq_vits_checkpoint -from TTS.tts.utils.helpers import generate_path, maximum_path, rand_segments, segment, sequence_mask +from TTS.tts.utils.helpers import generate_path, rand_segments, segment, sequence_mask from TTS.tts.utils.languages import LanguageManager from TTS.tts.utils.speakers import SpeakerManager from TTS.tts.utils.synthesis import synthesis from TTS.tts.utils.text.characters import BaseCharacters, BaseVocabulary, _characters, _pad, _phonemes, _punctuations from TTS.tts.utils.text.tokenizer import TTSTokenizer from TTS.tts.utils.visual import plot_alignment -from TTS.utils.io import load_fsspec +from TTS.utils.audio.torch_transforms import spec_to_mel, wav_to_mel, wav_to_spec from TTS.utils.samplers import BucketBatchSampler from TTS.vocoder.models.hifigan_generator import HifiganGenerator from TTS.vocoder.utils.generic_utils import plot_results +logger = logging.getLogger(__name__) + ############################## # IO / Feature extraction ############################## -# pylint: disable=global-statement -hann_window = {} -mel_basis = {} - @torch.no_grad() def weights_reset(m: nn.Module): @@ -75,139 +74,6 @@ def load_audio(file_path): return x, sr -def _amp_to_db(x, C=1, clip_val=1e-5): - return torch.log(torch.clamp(x, min=clip_val) * C) - - -def _db_to_amp(x, C=1): - return torch.exp(x) / C - - -def amp_to_db(magnitudes): - output = _amp_to_db(magnitudes) - return output - - -def db_to_amp(magnitudes): - output = _db_to_amp(magnitudes) - return output - - -def wav_to_spec(y, n_fft, hop_length, win_length, center=False): - """ - Args Shapes: - - y : :math:`[B, 1, T]` - - Return Shapes: - - spec : :math:`[B,C,T]` - """ - y = y.squeeze(1) - - if torch.min(y) < -1.0: - print("min value is ", torch.min(y)) - if torch.max(y) > 1.0: - print("max value is ", torch.max(y)) - - global hann_window - dtype_device = str(y.dtype) + "_" + str(y.device) - wnsize_dtype_device = str(win_length) + "_" + dtype_device - if wnsize_dtype_device not in hann_window: - hann_window[wnsize_dtype_device] = torch.hann_window(win_length).to(dtype=y.dtype, device=y.device) - - y = torch.nn.functional.pad( - y.unsqueeze(1), - (int((n_fft - hop_length) / 2), int((n_fft - hop_length) / 2)), - mode="reflect", - ) - y = y.squeeze(1) - - spec = torch.stft( - y, - n_fft, - hop_length=hop_length, - win_length=win_length, - window=hann_window[wnsize_dtype_device], - center=center, - pad_mode="reflect", - normalized=False, - onesided=True, - return_complex=False, - ) - - spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) - return spec - - -def spec_to_mel(spec, n_fft, num_mels, sample_rate, fmin, fmax): - """ - Args Shapes: - - spec : :math:`[B,C,T]` - - Return Shapes: - - mel : :math:`[B,C,T]` - """ - global mel_basis - dtype_device = str(spec.dtype) + "_" + str(spec.device) - fmax_dtype_device = str(fmax) + "_" + dtype_device - if fmax_dtype_device not in mel_basis: - mel = librosa_mel_fn(sr=sample_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax) - mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device) - mel = torch.matmul(mel_basis[fmax_dtype_device], spec) - mel = amp_to_db(mel) - return mel - - -def wav_to_mel(y, n_fft, num_mels, sample_rate, hop_length, win_length, fmin, fmax, center=False): - """ - Args Shapes: - - y : :math:`[B, 1, T]` - - Return Shapes: - - spec : :math:`[B,C,T]` - """ - y = y.squeeze(1) - - if torch.min(y) < -1.0: - print("min value is ", torch.min(y)) - if torch.max(y) > 1.0: - print("max value is ", torch.max(y)) - - global mel_basis, hann_window - dtype_device = str(y.dtype) + "_" + str(y.device) - fmax_dtype_device = str(fmax) + "_" + dtype_device - wnsize_dtype_device = str(win_length) + "_" + dtype_device - if fmax_dtype_device not in mel_basis: - mel = librosa_mel_fn(sr=sample_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax) - mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device) - if wnsize_dtype_device not in hann_window: - hann_window[wnsize_dtype_device] = torch.hann_window(win_length).to(dtype=y.dtype, device=y.device) - - y = torch.nn.functional.pad( - y.unsqueeze(1), - (int((n_fft - hop_length) / 2), int((n_fft - hop_length) / 2)), - mode="reflect", - ) - y = y.squeeze(1) - - spec = torch.stft( - y, - n_fft, - hop_length=hop_length, - win_length=win_length, - window=hann_window[wnsize_dtype_device], - center=center, - pad_mode="reflect", - normalized=False, - onesided=True, - return_complex=False, - ) - - spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) - spec = torch.matmul(mel_basis[fmax_dtype_device], spec) - spec = amp_to_db(spec) - return spec - - ############################# # CONFIGS ############################# @@ -229,30 +95,6 @@ class VitsAudioConfig(Coqpit): ############################## -def get_attribute_balancer_weights(items: list, attr_name: str, multi_dict: dict = None): - """Create inverse frequency weights for balancing the dataset. - Use `multi_dict` to scale relative weights.""" - attr_names_samples = np.array([item[attr_name] for item in items]) - unique_attr_names = np.unique(attr_names_samples).tolist() - attr_idx = [unique_attr_names.index(l) for l in attr_names_samples] - attr_count = np.array([len(np.where(attr_names_samples == l)[0]) for l in unique_attr_names]) - weight_attr = 1.0 / attr_count - dataset_samples_weight = np.array([weight_attr[l] for l in attr_idx]) - dataset_samples_weight = dataset_samples_weight / np.linalg.norm(dataset_samples_weight) - if multi_dict is not None: - # check if all keys are in the multi_dict - for k in multi_dict: - assert k in unique_attr_names, f"{k} not in {unique_attr_names}" - # scale weights - multiplier_samples = np.array([multi_dict.get(item[attr_name], 1.0) for item in items]) - dataset_samples_weight *= multiplier_samples - return ( - torch.from_numpy(dataset_samples_weight).float(), - unique_attr_names, - np.unique(dataset_samples_weight).tolist(), - ) - - class VitsDataset(TTSDataset): def __init__(self, model_args, *args, **kwargs): super().__init__(*args, **kwargs) @@ -760,7 +602,7 @@ def init_multispeaker(self, config: Coqpit): ) self.speaker_manager.encoder.eval() - print(" > External Speaker Encoder Loaded !!") + logger.info("External Speaker Encoder Loaded !!") if ( hasattr(self.speaker_manager.encoder, "audio_config") @@ -774,7 +616,7 @@ def init_multispeaker(self, config: Coqpit): def _init_speaker_embedding(self): # pylint: disable=attribute-defined-outside-init if self.num_speakers > 0: - print(" > initialization of speaker-embedding layers.") + logger.info("Initialization of speaker-embedding layers.") self.embedded_speaker_dim = self.args.speaker_embedding_channels self.emb_g = nn.Embedding(self.num_speakers, self.embedded_speaker_dim) @@ -794,7 +636,7 @@ def init_multilingual(self, config: Coqpit): self.language_manager = LanguageManager(language_ids_file_path=config.language_ids_file) if self.args.use_language_embedding and self.language_manager: - print(" > initialization of language-embedding layers.") + logger.info("Initialization of language-embedding layers.") self.num_languages = self.language_manager.num_languages self.embedded_language_dim = self.args.embedded_language_dim self.emb_l = nn.Embedding(self.num_languages, self.embedded_language_dim) @@ -829,7 +671,7 @@ def on_init_end(self, trainer): # pylint: disable=W0613 for key, value in after_dict.items(): if value == before_dict[key]: raise RuntimeError(" [!] The weights of Duration Predictor was not reinit check it !") - print(" > Duration Predictor was reinit.") + logger.info("Duration Predictor was reinit.") if self.args.reinit_text_encoder: before_dict = get_module_weights_sum(self.text_encoder) @@ -839,7 +681,7 @@ def on_init_end(self, trainer): # pylint: disable=W0613 for key, value in after_dict.items(): if value == before_dict[key]: raise RuntimeError(" [!] The weights of Text Encoder was not reinit check it !") - print(" > Text Encoder was reinit.") + logger.info("Text Encoder was reinit.") def get_aux_input(self, aux_input: Dict): sid, g, lid, _ = self._set_cond_input(aux_input) @@ -1233,7 +1075,7 @@ def train_step(self, batch: dict, criterion: nn.Module, optimizer_idx: int) -> T Args: batch (Dict): Input tensors. criterion (nn.Module): Loss layer designed for the model. - optimizer_idx (int): Index of optimizer to use. 0 for the generator and 1 for the discriminator networks. + optimizer_idx (int): Index of optimizer to use. 0 for the discriminator and 1 for the generator networks. Returns: Tuple[Dict, Dict]: Model ouputs and computed losses. @@ -1270,7 +1112,7 @@ def train_step(self, batch: dict, criterion: nn.Module, optimizer_idx: int) -> T ) # compute loss - with autocast(enabled=False): # use float32 for the criterion + with torch.autocast("cuda", enabled=False): # use float32 for the criterion loss_dict = criterion[optimizer_idx]( scores_disc_real, scores_disc_fake, @@ -1281,7 +1123,7 @@ def train_step(self, batch: dict, criterion: nn.Module, optimizer_idx: int) -> T mel = batch["mel"] # compute melspec segment - with autocast(enabled=False): + with torch.autocast("cuda", enabled=False): if self.args.encoder_sample_rate: spec_segment_size = self.spec_segment_size * int(self.interpolate_factor) else: @@ -1308,7 +1150,7 @@ def train_step(self, batch: dict, criterion: nn.Module, optimizer_idx: int) -> T ) # compute losses - with autocast(enabled=False): # use float32 for the criterion + with torch.autocast("cuda", enabled=False): # use float32 for the criterion loss_dict = criterion[optimizer_idx]( mel_slice_hat=mel_slice.float(), mel_slice=mel_slice_hat.float(), @@ -1433,7 +1275,7 @@ def test_run(self, assets) -> Tuple[Dict, Dict]: Returns: Tuple[Dict, Dict]: Test figures and audios to be projected to Tensorboard. """ - print(" | > Synthesizing test sentences.") + logger.info("Synthesizing test sentences.") test_audios = {} test_figures = {} test_sentences = self.config.test_sentences @@ -1550,14 +1392,14 @@ def get_sampler(self, config: Coqpit, dataset: TTSDataset, num_gpus=1, is_eval=F data_items = dataset.samples if getattr(config, "use_weighted_sampler", False): for attr_name, alpha in config.weighted_sampler_attrs.items(): - print(f" > Using weighted sampler for attribute '{attr_name}' with alpha '{alpha}'") + logger.info("Using weighted sampler for attribute '%s' with alpha %.3f", attr_name, alpha) multi_dict = config.weighted_sampler_multipliers.get(attr_name, None) - print(multi_dict) + logger.info(multi_dict) weights, attr_names, attr_weights = get_attribute_balancer_weights( attr_name=attr_name, items=data_items, multi_dict=multi_dict ) weights = weights * alpha - print(f" > Attribute weights for '{attr_names}' \n | > {attr_weights}") + logger.info("Attribute weights for '%s' \n | > %s", attr_names, attr_weights) # input_audio_lenghts = [os.path.getsize(x["audio_file"]) for x in data_items] @@ -1605,7 +1447,6 @@ def get_data_loader( max_audio_len=config.max_audio_len, phoneme_cache_path=config.phoneme_cache_path, precompute_num_workers=config.precompute_num_workers, - verbose=verbose, tokenizer=self.tokenizer, start_by_longest=config.start_by_longest, ) @@ -1651,13 +1492,16 @@ def get_data_loader( def get_optimizer(self) -> List: """Initiate and return the GAN optimizers based on the config parameters. - It returnes 2 optimizers in a list. First one is for the generator and the second one is for the discriminator. + + It returns 2 optimizers in a list. First one is for the discriminator + and the second one is for the generator. + Returns: List: optimizers. """ - # select generator parameters optimizer0 = get_optimizer(self.config.optimizer, self.config.optimizer_params, self.config.lr_disc, self.disc) + # select generator parameters gen_parameters = chain(params for k, params in self.named_parameters() if not k.startswith("disc.")) optimizer1 = get_optimizer( self.config.optimizer, self.config.optimizer_params, self.config.lr_gen, parameters=gen_parameters @@ -1712,7 +1556,7 @@ def load_checkpoint( # handle fine-tuning from a checkpoint with additional speakers if hasattr(self, "emb_g") and state["model"]["emb_g.weight"].shape != self.emb_g.weight.shape: num_new_speakers = self.emb_g.weight.shape[0] - state["model"]["emb_g.weight"].shape[0] - print(f" > Loading checkpoint with {num_new_speakers} additional speakers.") + logger.info("Loading checkpoint with %d additional speakers.", num_new_speakers) emb_g = state["model"]["emb_g.weight"] new_row = torch.randn(num_new_speakers, emb_g.shape[1]) emb_g = torch.cat([emb_g, new_row], axis=0) @@ -1769,7 +1613,7 @@ def load_fairseq_checkpoint( assert not self.training @staticmethod - def init_from_config(config: "VitsConfig", samples: Union[List[List], List[Dict]] = None, verbose=True): + def init_from_config(config: "VitsConfig", samples: Union[List[List], List[Dict]] = None): """Initiate model from config Args: @@ -1792,7 +1636,7 @@ def init_from_config(config: "VitsConfig", samples: Union[List[List], List[Dict] upsample_rate == effective_hop_length ), f" [!] Product of upsample rates must be equal to the hop length - {upsample_rate} vs {effective_hop_length}" - ap = AudioProcessor.init_from_config(config, verbose=verbose) + ap = AudioProcessor.init_from_config(config) tokenizer, new_config = TTSTokenizer.init_from_config(config) speaker_manager = SpeakerManager.init_from_config(config, samples) language_manager = LanguageManager.init_from_config(config) @@ -1880,16 +1724,18 @@ def onnx_inference(text, text_lengths, scales, sid=None, langid=None): self.forward = _forward if training: self.train() - if not disc is None: + if disc is not None: self.disc = disc def load_onnx(self, model_path: str, cuda=False): import onnxruntime as ort providers = [ - "CPUExecutionProvider" - if cuda is False - else ("CUDAExecutionProvider", {"cudnn_conv_algo_search": "DEFAULT"}) + ( + "CPUExecutionProvider" + if cuda is False + else ("CUDAExecutionProvider", {"cudnn_conv_algo_search": "DEFAULT"}) + ) ] sess_options = ort.SessionOptions() self.onnx_sess = ort.InferenceSession( @@ -1914,9 +1760,9 @@ def inference_onnx(self, x, x_lengths=None, speaker_id=None, language_id=None): dtype=np.float32, ) input_params = {"input": x, "input_lengths": x_lengths, "scales": scales} - if not speaker_id is None: + if speaker_id is not None: input_params["sid"] = torch.tensor([speaker_id]).cpu().numpy() - if not language_id is None: + if language_id is not None: input_params["langid"] = torch.tensor([language_id]).cpu().numpy() audio = self.onnx_sess.run( @@ -1948,8 +1794,7 @@ def __init__( def _create_vocab(self): self._vocab = [self._pad] + list(self._punctuations) + list(self._characters) + [self._blank] self._char_to_id = {char: idx for idx, char in enumerate(self.vocab)} - # pylint: disable=unnecessary-comprehension - self._id_to_char = {idx: char for idx, char in enumerate(self.vocab)} + self._id_to_char = dict(enumerate(self.vocab)) @staticmethod def init_from_config(config: Coqpit): @@ -1996,4 +1841,4 @@ def vocab(self, vocab_file): self.blank = self._vocab[0] self.pad = " " self._char_to_id = {s: i for i, s in enumerate(self._vocab)} # pylint: disable=unnecessary-comprehension - self._id_to_char = {i: s for i, s in enumerate(self._vocab)} # pylint: disable=unnecessary-comprehension + self._id_to_char = dict(enumerate(self._vocab)) diff --git a/TTS/tts/models/xtts.py b/TTS/tts/models/xtts.py index a109c6e78b..38091d7cff 100644 --- a/TTS/tts/models/xtts.py +++ b/TTS/tts/models/xtts.py @@ -1,19 +1,25 @@ +import logging import os from dataclasses import dataclass +from pathlib import Path +from typing import Optional import librosa import torch import torch.nn.functional as F import torchaudio from coqpit import Coqpit +from trainer.io import load_fsspec from TTS.tts.layers.xtts.gpt import GPT from TTS.tts.layers.xtts.hifigan_decoder import HifiDecoder from TTS.tts.layers.xtts.stream_generator import init_stream_support from TTS.tts.layers.xtts.tokenizer import VoiceBpeTokenizer, split_sentence -from TTS.tts.layers.xtts.xtts_manager import SpeakerManager, LanguageManager +from TTS.tts.layers.xtts.xtts_manager import LanguageManager, SpeakerManager from TTS.tts.models.base_tts import BaseTTS -from TTS.utils.io import load_fsspec +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 + +logger = logging.getLogger(__name__) init_stream_support() @@ -61,7 +67,7 @@ def wav_to_mel_cloning( mel = mel_stft(wav) mel = torch.log(torch.clamp(mel, min=1e-5)) if mel_norms is None: - mel_norms = torch.load(mel_norms_file, map_location=device) + mel_norms = torch.load(mel_norms_file, map_location=device, weights_only=is_pytorch_at_least_2_4()) mel = mel / mel_norms.unsqueeze(0).unsqueeze(-1) return mel @@ -82,31 +88,12 @@ def load_audio(audiopath, sampling_rate): # Check some assumptions about audio range. This should be automatically fixed in load_wav_to_torch, but might not be in some edge cases, where we should squawk. # '10' is arbitrarily chosen since it seems like audio will often "overdrive" the [-1,1] bounds. if torch.any(audio > 10) or not torch.any(audio < 0): - print(f"Error with {audiopath}. Max={audio.max()} min={audio.min()}") + logger.error("Error with %s. Max=%.2f min=%.2f", audiopath, audio.max(), audio.min()) # clip audio invalid values audio.clip_(-1, 1) return audio -def pad_or_truncate(t, length): - """ - Ensure a given tensor t has a specified sequence length by either padding it with zeros or clipping it. - - Args: - t (torch.Tensor): The input tensor to be padded or truncated. - length (int): The desired length of the tensor. - - Returns: - torch.Tensor: The padded or truncated tensor. - """ - tp = t[..., :length] - if t.shape[-1] == length: - tp = t - elif t.shape[-1] < length: - tp = F.pad(t, (0, length - t.shape[-1])) - return tp - - @dataclass class XttsAudioConfig(Coqpit): """ @@ -115,10 +102,12 @@ class XttsAudioConfig(Coqpit): Args: sample_rate (int): The sample rate in which the GPT operates. output_sample_rate (int): The sample rate of the output audio waveform. + dvae_sample_rate (int): The sample rate of the DVAE """ sample_rate: int = 22050 output_sample_rate: int = 24000 + dvae_sample_rate: int = 22050 @dataclass @@ -189,7 +178,7 @@ class XttsArgs(Coqpit): class Xtts(BaseTTS): - """ⓍTTS model implementation. + """XTTS model implementation. ❗ Currently it only supports inference. @@ -197,7 +186,7 @@ class Xtts(BaseTTS): >>> from TTS.tts.configs.xtts_config import XttsConfig >>> from TTS.tts.models.xtts import Xtts >>> config = XttsConfig() - >>> model = Xtts.inif_from_config(config) + >>> model = Xtts.init_from_config(config) >>> model.load_checkpoint(config, checkpoint_dir="paths/to/models_dir/", eval=True) """ @@ -274,7 +263,7 @@ def get_gpt_cond_latents(self, audio, sr, length: int = 30, chunk_length: int = for i in range(0, audio.shape[1], 22050 * chunk_length): audio_chunk = audio[:, i : i + 22050 * chunk_length] - # if the chunk is too short ignore it + # if the chunk is too short ignore it if audio_chunk.size(-1) < 22050 * 0.33: continue @@ -410,12 +399,14 @@ def synthesize(self, text, config, speaker_wav, language, speaker_id=None, speed if speaker_id is not None: gpt_cond_latent, speaker_embedding = self.speaker_manager.speakers[speaker_id].values() return self.inference(text, language, gpt_cond_latent, speaker_embedding, speed=speed, **settings) - settings.update({ + settings.update( + { "gpt_cond_len": config.gpt_cond_len, "gpt_cond_chunk_len": config.gpt_cond_chunk_len, "max_ref_len": config.max_ref_len, "sound_norm_refs": config.sound_norm_refs, - }) + } + ) return self.full_inference(text, speaker_wav, language, speed=speed, **settings) @torch.inference_mode() @@ -470,7 +461,7 @@ def full_inference( gpt_cond_chunk_len: (int) Chunk length used for cloning. It must be <= `gpt_cond_len`. If gpt_cond_len == gpt_cond_chunk_len, no chunking. Defaults to 6 seconds. - hf_generate_kwargs: (**kwargs) The huggingface Transformers generate API is used for the autoregressive + hf_generate_kwargs: (`**kwargs`) The huggingface Transformers generate API is used for the autoregressive transformer. Extra keyword args fed to this function get forwarded directly to that API. Documentation here: https://huggingface.co/docs/transformers/internal/generation_utils @@ -660,6 +651,7 @@ def inference_stream( repetition_penalty=float(repetition_penalty), output_attentions=False, output_hidden_states=True, + return_dict_in_generate=True, **hf_generate_kwargs, ) @@ -692,12 +684,12 @@ def inference_stream( def forward(self): raise NotImplementedError( - "XTTS has a dedicated trainer, please check the XTTS docs: https://tts.readthedocs.io/en/dev/models/xtts.html#training" + "XTTS has a dedicated trainer, please check the XTTS docs: https://coqui-tts.readthedocs.io/en/latest/models/xtts.html#training" ) def eval_step(self): raise NotImplementedError( - "XTTS has a dedicated trainer, please check the XTTS docs: https://tts.readthedocs.io/en/dev/models/xtts.html#training" + "XTTS has a dedicated trainer, please check the XTTS docs: https://coqui-tts.readthedocs.io/en/latest/models/xtts.html#training" ) @staticmethod @@ -729,14 +721,14 @@ def get_compatible_checkpoint_state_dict(self, model_path): def load_checkpoint( self, - config, - checkpoint_dir=None, - checkpoint_path=None, - vocab_path=None, - eval=True, - strict=True, - use_deepspeed=False, - speaker_file_path=None, + config: "XttsConfig", + checkpoint_dir: Optional[str] = None, + checkpoint_path: Optional[str] = None, + vocab_path: Optional[str] = None, + eval: bool = True, + strict: bool = True, + use_deepspeed: bool = False, + speaker_file_path: Optional[str] = None, ): """ Loads a checkpoint from disk and initializes the model's state and tokenizer. @@ -752,9 +744,15 @@ def load_checkpoint( Returns: None """ - + if checkpoint_dir is not None and Path(checkpoint_dir).is_file(): + msg = f"You passed a file to `checkpoint_dir=`. Use `checkpoint_path={checkpoint_dir}` instead." + raise ValueError(msg) model_path = checkpoint_path or os.path.join(checkpoint_dir, "model.pth") - vocab_path = vocab_path or os.path.join(checkpoint_dir, "vocab.json") + if vocab_path is None: + if checkpoint_dir is not None and (Path(checkpoint_dir) / "vocab.json").is_file(): + vocab_path = str(Path(checkpoint_dir) / "vocab.json") + else: + vocab_path = config.model_args.tokenizer_file if speaker_file_path is None and checkpoint_dir is not None: speaker_file_path = os.path.join(checkpoint_dir, "speakers_xtts.pth") @@ -766,6 +764,12 @@ def load_checkpoint( if os.path.exists(vocab_path): self.tokenizer = VoiceBpeTokenizer(vocab_file=vocab_path) + else: + msg = ( + f"`vocab.json` file not found in `{checkpoint_dir}`. Move the file there or " + "specify alternative path in `model_args.tokenizer_file` in `config.json`" + ) + raise FileNotFoundError(msg) self.init_models() @@ -786,5 +790,5 @@ def load_checkpoint( def train_step(self): raise NotImplementedError( - "XTTS has a dedicated trainer, please check the XTTS docs: https://tts.readthedocs.io/en/dev/models/xtts.html#training" + "XTTS has a dedicated trainer, please check the XTTS docs: https://coqui-tts.readthedocs.io/en/latest/models/xtts.html#training" ) diff --git a/TTS/tts/utils/assets/tortoise/tokenizer.json b/TTS/tts/utils/assets/tortoise/tokenizer.json index a128f27305..c2fb44a729 100644 --- a/TTS/tts/utils/assets/tortoise/tokenizer.json +++ b/TTS/tts/utils/assets/tortoise/tokenizer.json @@ -1 +1 @@ -{"version":"1.0","truncation":null,"padding":null,"added_tokens":[{"id":0,"special":true,"content":"[STOP]","single_word":false,"lstrip":false,"rstrip":false,"normalized":false},{"id":1,"special":true,"content":"[UNK]","single_word":false,"lstrip":false,"rstrip":false,"normalized":false},{"id":2,"special":true,"content":"[SPACE]","single_word":false,"lstrip":false,"rstrip":false,"normalized":false}],"normalizer":null,"pre_tokenizer":{"type":"Whitespace"},"post_processor":null,"decoder":null,"model":{"type":"BPE","dropout":null,"unk_token":"[UNK]","continuing_subword_prefix":null,"end_of_word_suffix":null,"fuse_unk":false,"vocab":{"[STOP]":0,"[UNK]":1,"[SPACE]":2,"!":3,"'":4,"(":5,")":6,",":7,"-":8,".":9,"/":10,":":11,";":12,"?":13,"a":14,"b":15,"c":16,"d":17,"e":18,"f":19,"g":20,"h":21,"i":22,"j":23,"k":24,"l":25,"m":26,"n":27,"o":28,"p":29,"q":30,"r":31,"s":32,"t":33,"u":34,"v":35,"w":36,"x":37,"y":38,"z":39,"th":40,"in":41,"the":42,"an":43,"er":44,"ou":45,"re":46,"on":47,"at":48,"ed":49,"en":50,"to":51,"ing":52,"and":53,"is":54,"as":55,"al":56,"or":57,"of":58,"ar":59,"it":60,"es":61,"he":62,"st":63,"le":64,"om":65,"se":66,"be":67,"ad":68,"ow":69,"ly":70,"ch":71,"wh":72,"that":73,"you":74,"li":75,"ve":76,"ac":77,"ti":78,"ld":79,"me":80,"was":81,"gh":82,"id":83,"ll":84,"wi":85,"ent":86,"for":87,"ay":88,"ro":89,"ver":90,"ic":91,"her":92,"ke":93,"his":94,"no":95,"ut":96,"un":97,"ir":98,"lo":99,"we":100,"ri":101,"ha":102,"with":103,"ght":104,"out":105,"im":106,"ion":107,"all":108,"ab":109,"one":110,"ne":111,"ge":112,"ould":113,"ter":114,"mo":115,"had":116,"ce":117,"she":118,"go":119,"sh":120,"ur":121,"am":122,"so":123,"pe":124,"my":125,"de":126,"are":127,"but":128,"ome":129,"fr":130,"ther":131,"fe":132,"su":133,"do":134,"con":135,"te":136,"ain":137,"ere":138,"po":139,"if":140,"they":141,"us":142,"ag":143,"tr":144,"now":145,"oun":146,"this":147,"have":148,"not":149,"sa":150,"il":151,"up":152,"thing":153,"from":154,"ap":155,"him":156,"ack":157,"ation":158,"ant":159,"our":160,"op":161,"like":162,"ust":163,"ess":164,"bo":165,"ok":166,"ul":167,"ind":168,"ex":169,"com":170,"some":171,"there":172,"ers":173,"co":174,"res":175,"man":176,"ard":177,"pl":178,"wor":179,"way":180,"tion":181,"fo":182,"ca":183,"were":184,"by":185,"ate":186,"pro":187,"ted":188,"ound":189,"own":190,"would":191,"ts":192,"what":193,"qu":194,"ally":195,"ight":196,"ck":197,"gr":198,"when":199,"ven":200,"can":201,"ough":202,"ine":203,"end":204,"per":205,"ous":206,"od":207,"ide":208,"know":209,"ty":210,"very":211,"si":212,"ak":213,"who":214,"about":215,"ill":216,"them":217,"est":218,"red":219,"ye":220,"could":221,"ong":222,"your":223,"their":224,"em":225,"just":226,"other":227,"into":228,"any":229,"whi":230,"um":231,"tw":232,"ast":233,"der":234,"did":235,"ie":236,"been":237,"ace":238,"ink":239,"ity":240,"back":241,"ting":242,"br":243,"more":244,"ake":245,"pp":246,"then":247,"sp":248,"el":249,"use":250,"bl":251,"said":252,"over":253,"get":254},"merges":["t h","i n","th e","a n","e r","o u","r e","o n","a t","e d","e n","t o","in g","an d","i s","a s","a l","o r","o f","a r","i t","e s","h e","s t","l e","o m","s e","b e","a d","o w","l y","c h","w h","th at","y ou","l i","v e","a c","t i","l d","m e","w as","g h","i d","l l","w i","en t","f or","a y","r o","v er","i c","h er","k e","h is","n o","u t","u n","i r","l o","w e","r i","h a","wi th","gh t","ou t","i m","i on","al l","a b","on e","n e","g e","ou ld","t er","m o","h ad","c e","s he","g o","s h","u r","a m","s o","p e","m y","d e","a re","b ut","om e","f r","the r","f e","s u","d o","c on","t e","a in","er e","p o","i f","the y","u s","a g","t r","n ow","ou n","th is","ha ve","no t","s a","i l","u p","th ing","fr om","a p","h im","ac k","at ion","an t","ou r","o p","li ke","u st","es s","b o","o k","u l","in d","e x","c om","s ome","the re","er s","c o","re s","m an","ar d","p l","w or","w ay","ti on","f o","c a","w ere","b y","at e","p ro","t ed","oun d","ow n","w ould","t s","wh at","q u","al ly","i ght","c k","g r","wh en","v en","c an","ou gh","in e","en d","p er","ou s","o d","id e","k now","t y","ver y","s i","a k","wh o","ab out","i ll","the m","es t","re d","y e","c ould","on g","you r","the ir","e m","j ust","o ther","in to","an y","wh i","u m","t w","as t","d er","d id","i e","be en","ac e","in k","it y","b ack","t ing","b r","mo re","a ke","p p","the n","s p","e l","u se","b l","sa id","o ver","ge t"]}} \ No newline at end of file +{"version":"1.0","truncation":null,"padding":null,"added_tokens":[{"id":0,"special":true,"content":"[STOP]","single_word":false,"lstrip":false,"rstrip":false,"normalized":false},{"id":1,"special":true,"content":"[UNK]","single_word":false,"lstrip":false,"rstrip":false,"normalized":false},{"id":2,"special":true,"content":"[SPACE]","single_word":false,"lstrip":false,"rstrip":false,"normalized":false}],"normalizer":null,"pre_tokenizer":{"type":"Whitespace"},"post_processor":null,"decoder":null,"model":{"type":"BPE","dropout":null,"unk_token":"[UNK]","continuing_subword_prefix":null,"end_of_word_suffix":null,"fuse_unk":false,"vocab":{"[STOP]":0,"[UNK]":1,"[SPACE]":2,"!":3,"'":4,"(":5,")":6,",":7,"-":8,".":9,"/":10,":":11,";":12,"?":13,"a":14,"b":15,"c":16,"d":17,"e":18,"f":19,"g":20,"h":21,"i":22,"j":23,"k":24,"l":25,"m":26,"n":27,"o":28,"p":29,"q":30,"r":31,"s":32,"t":33,"u":34,"v":35,"w":36,"x":37,"y":38,"z":39,"th":40,"in":41,"the":42,"an":43,"er":44,"ou":45,"re":46,"on":47,"at":48,"ed":49,"en":50,"to":51,"ing":52,"and":53,"is":54,"as":55,"al":56,"or":57,"of":58,"ar":59,"it":60,"es":61,"he":62,"st":63,"le":64,"om":65,"se":66,"be":67,"ad":68,"ow":69,"ly":70,"ch":71,"wh":72,"that":73,"you":74,"li":75,"ve":76,"ac":77,"ti":78,"ld":79,"me":80,"was":81,"gh":82,"id":83,"ll":84,"wi":85,"ent":86,"for":87,"ay":88,"ro":89,"ver":90,"ic":91,"her":92,"ke":93,"his":94,"no":95,"ut":96,"un":97,"ir":98,"lo":99,"we":100,"ri":101,"ha":102,"with":103,"ght":104,"out":105,"im":106,"ion":107,"all":108,"ab":109,"one":110,"ne":111,"ge":112,"ould":113,"ter":114,"mo":115,"had":116,"ce":117,"she":118,"go":119,"sh":120,"ur":121,"am":122,"so":123,"pe":124,"my":125,"de":126,"are":127,"but":128,"ome":129,"fr":130,"ther":131,"fe":132,"su":133,"do":134,"con":135,"te":136,"ain":137,"ere":138,"po":139,"if":140,"they":141,"us":142,"ag":143,"tr":144,"now":145,"oun":146,"this":147,"have":148,"not":149,"sa":150,"il":151,"up":152,"thing":153,"from":154,"ap":155,"him":156,"ack":157,"ation":158,"ant":159,"our":160,"op":161,"like":162,"ust":163,"ess":164,"bo":165,"ok":166,"ul":167,"ind":168,"ex":169,"com":170,"some":171,"there":172,"ers":173,"co":174,"res":175,"man":176,"ard":177,"pl":178,"wor":179,"way":180,"tion":181,"fo":182,"ca":183,"were":184,"by":185,"ate":186,"pro":187,"ted":188,"ound":189,"own":190,"would":191,"ts":192,"what":193,"qu":194,"ally":195,"ight":196,"ck":197,"gr":198,"when":199,"ven":200,"can":201,"ough":202,"ine":203,"end":204,"per":205,"ous":206,"od":207,"ide":208,"know":209,"ty":210,"very":211,"si":212,"ak":213,"who":214,"about":215,"ill":216,"them":217,"est":218,"red":219,"ye":220,"could":221,"ong":222,"your":223,"their":224,"em":225,"just":226,"other":227,"into":228,"any":229,"whi":230,"um":231,"tw":232,"ast":233,"der":234,"did":235,"ie":236,"been":237,"ace":238,"ink":239,"ity":240,"back":241,"ting":242,"br":243,"more":244,"ake":245,"pp":246,"then":247,"sp":248,"el":249,"use":250,"bl":251,"said":252,"over":253,"get":254},"merges":["t h","i n","th e","a n","e r","o u","r e","o n","a t","e d","e n","t o","in g","an d","i s","a s","a l","o r","o f","a r","i t","e s","h e","s t","l e","o m","s e","b e","a d","o w","l y","c h","w h","th at","y ou","l i","v e","a c","t i","l d","m e","w as","g h","i d","l l","w i","en t","f or","a y","r o","v er","i c","h er","k e","h is","n o","u t","u n","i r","l o","w e","r i","h a","wi th","gh t","ou t","i m","i on","al l","a b","on e","n e","g e","ou ld","t er","m o","h ad","c e","s he","g o","s h","u r","a m","s o","p e","m y","d e","a re","b ut","om e","f r","the r","f e","s u","d o","c on","t e","a in","er e","p o","i f","the y","u s","a g","t r","n ow","ou n","th is","ha ve","no t","s a","i l","u p","th ing","fr om","a p","h im","ac k","at ion","an t","ou r","o p","li ke","u st","es s","b o","o k","u l","in d","e x","c om","s ome","the re","er s","c o","re s","m an","ar d","p l","w or","w ay","ti on","f o","c a","w ere","b y","at e","p ro","t ed","oun d","ow n","w ould","t s","wh at","q u","al ly","i ght","c k","g r","wh en","v en","c an","ou gh","in e","en d","p er","ou s","o d","id e","k now","t y","ver y","s i","a k","wh o","ab out","i ll","the m","es t","re d","y e","c ould","on g","you r","the ir","e m","j ust","o ther","in to","an y","wh i","u m","t w","as t","d er","d id","i e","be en","ac e","in k","it y","b ack","t ing","b r","mo re","a ke","p p","the n","s p","e l","u se","b l","sa id","o ver","ge t"]}} diff --git a/TTS/tts/utils/fairseq.py b/TTS/tts/utils/fairseq.py index 3d8eec2b4e..20907a0532 100644 --- a/TTS/tts/utils/fairseq.py +++ b/TTS/tts/utils/fairseq.py @@ -1,8 +1,10 @@ import torch +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 + def rehash_fairseq_vits_checkpoint(checkpoint_file): - chk = torch.load(checkpoint_file, map_location=torch.device("cpu"))["model"] + chk = torch.load(checkpoint_file, map_location=torch.device("cpu"), weights_only=is_pytorch_at_least_2_4())["model"] new_chk = {} for k, v in chk.items(): if "enc_p." in k: diff --git a/TTS/tts/utils/helpers.py b/TTS/tts/utils/helpers.py index 7b37201f84..ff10f751f2 100644 --- a/TTS/tts/utils/helpers.py +++ b/TTS/tts/utils/helpers.py @@ -1,15 +1,10 @@ +from typing import Optional + import numpy as np import torch from scipy.stats import betabinom from torch.nn import functional as F -try: - from TTS.tts.utils.monotonic_align.core import maximum_path_c - - CYTHON = True -except ModuleNotFoundError: - CYTHON = False - class StandardScaler: """StandardScaler for mean-scale normalization with the given mean and scale values.""" @@ -40,7 +35,7 @@ def inverse_transform(self, X): # from https://gist.github.com/jihunchoi/f1434a77df9db1bb337417854b398df1 -def sequence_mask(sequence_length, max_len=None): +def sequence_mask(sequence_length: torch.Tensor, max_len: Optional[int] = None) -> torch.Tensor: """Create a sequence mask for filtering padding in a sequence tensor. Args: @@ -51,7 +46,7 @@ def sequence_mask(sequence_length, max_len=None): - mask: :math:`[B, T_max]` """ if max_len is None: - max_len = sequence_length.max() + max_len = int(sequence_length.max()) seq_range = torch.arange(max_len, dtype=sequence_length.dtype, device=sequence_length.device) # B x T_max return seq_range.unsqueeze(0) < sequence_length.unsqueeze(1) @@ -145,95 +140,80 @@ def average_over_durations(values, durs): return avg -def convert_pad_shape(pad_shape): +def convert_pad_shape(pad_shape: list[list]) -> list: l = pad_shape[::-1] - pad_shape = [item for sublist in l for item in sublist] - return pad_shape + return [item for sublist in l for item in sublist] -def generate_path(duration, mask): - """ +def generate_path(duration: torch.Tensor, mask: torch.Tensor) -> torch.Tensor: + """Generate alignment path based on the given segment durations. + Shapes: - duration: :math:`[B, T_en]` - mask: :math:'[B, T_en, T_de]` - path: :math:`[B, T_en, T_de]` """ b, t_x, t_y = mask.shape - cum_duration = torch.cumsum(duration, 1) + cum_duration = torch.cumsum(duration, dim=1) cum_duration_flat = cum_duration.view(b * t_x) path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype) path = path.view(b, t_x, t_y) path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1] - path = path * mask - return path + return path * mask -def maximum_path(value, mask): - if CYTHON: - return maximum_path_cython(value, mask) - return maximum_path_numpy(value, mask) +def generate_attention( + duration: torch.Tensor, x_mask: torch.Tensor, y_mask: Optional[torch.Tensor] = None +) -> torch.Tensor: + """Generate an attention map from the linear scale durations. + Args: + duration (Tensor): Linear scale durations. + x_mask (Tensor): Mask for the input (character) sequence. + y_mask (Tensor): Mask for the output (spectrogram) sequence. Compute it from the predicted durations + if None. Defaults to None. + + Shapes + - duration: :math:`(B, T_{en})` + - x_mask: :math:`(B, T_{en})` + - y_mask: :math:`(B, T_{de})` + """ + # compute decode mask from the durations + if y_mask is None: + y_lengths = duration.sum(dim=1).long() + y_lengths[y_lengths < 1] = 1 + y_mask = sequence_mask(y_lengths).unsqueeze(1).to(duration.dtype) + attn_mask = x_mask.unsqueeze(-1) * y_mask.unsqueeze(2) + return generate_path(duration, attn_mask.squeeze(1)).to(duration.dtype) + + +def expand_encoder_outputs( + x: torch.Tensor, duration: torch.Tensor, x_mask: torch.Tensor, y_lengths: torch.Tensor +) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Generate attention alignment map from durations and expand encoder outputs. -def maximum_path_cython(value, mask): - """Cython optimised version. Shapes: - - value: :math:`[B, T_en, T_de]` - - mask: :math:`[B, T_en, T_de]` - """ - value = value * mask - device = value.device - dtype = value.dtype - value = value.data.cpu().numpy().astype(np.float32) - path = np.zeros_like(value).astype(np.int32) - mask = mask.data.cpu().numpy() + - x: Encoder output :math:`(B, D_{en}, T_{en})` + - duration: :math:`(B, T_{en})` + - x_mask: :math:`(B, T_{en})` + - y_lengths: :math:`(B)` - t_x_max = mask.sum(1)[:, 0].astype(np.int32) - t_y_max = mask.sum(2)[:, 0].astype(np.int32) - maximum_path_c(path, value, t_x_max, t_y_max) - return torch.from_numpy(path).to(device=device, dtype=dtype) + Examples:: + encoder output: [a,b,c,d] + durations: [1, 3, 2, 1] -def maximum_path_numpy(value, mask, max_neg_val=None): - """ - Monotonic alignment search algorithm - Numpy-friendly version. It's about 4 times faster than torch version. - value: [b, t_x, t_y] - mask: [b, t_x, t_y] + expanded: [a, b, b, b, c, c, d] + attention map: [[0, 0, 0, 0, 0, 0, 1], + [0, 0, 0, 0, 1, 1, 0], + [0, 1, 1, 1, 0, 0, 0], + [1, 0, 0, 0, 0, 0, 0]] """ - if max_neg_val is None: - max_neg_val = -np.inf # Patch for Sphinx complaint - value = value * mask - - device = value.device - dtype = value.dtype - value = value.cpu().detach().numpy() - mask = mask.cpu().detach().numpy().astype(bool) - - b, t_x, t_y = value.shape - direction = np.zeros(value.shape, dtype=np.int64) - v = np.zeros((b, t_x), dtype=np.float32) - x_range = np.arange(t_x, dtype=np.float32).reshape(1, -1) - for j in range(t_y): - v0 = np.pad(v, [[0, 0], [1, 0]], mode="constant", constant_values=max_neg_val)[:, :-1] - v1 = v - max_mask = v1 >= v0 - v_max = np.where(max_mask, v1, v0) - direction[:, :, j] = max_mask - - index_mask = x_range <= j - v = np.where(index_mask, v_max + value[:, :, j], max_neg_val) - direction = np.where(mask, direction, 1) - - path = np.zeros(value.shape, dtype=np.float32) - index = mask[:, :, 0].sum(1).astype(np.int64) - 1 - index_range = np.arange(b) - for j in reversed(range(t_y)): - path[index_range, index, j] = 1 - index = index + direction[index_range, index, j] - 1 - path = path * mask.astype(np.float32) - path = torch.from_numpy(path).to(device=device, dtype=dtype) - return path + y_mask = sequence_mask(y_lengths).unsqueeze(1).to(x.dtype) + attn = generate_attention(duration, x_mask, y_mask) + x_expanded = torch.einsum("kmn, kjm -> kjn", [attn.float(), x]) + return x_expanded, attn, y_mask def beta_binomial_prior_distribution(phoneme_count, mel_count, scaling_factor=1.0): diff --git a/TTS/tts/utils/languages.py b/TTS/tts/utils/languages.py index 1e1836b32c..c72de2d4e6 100644 --- a/TTS/tts/utils/languages.py +++ b/TTS/tts/utils/languages.py @@ -1,5 +1,5 @@ import os -from typing import Any, Dict, List +from typing import Any, Dict, List, Optional, Union import fsspec import numpy as np @@ -27,8 +27,8 @@ class LanguageManager(BaseIDManager): def __init__( self, - language_ids_file_path: str = "", - config: Coqpit = None, + language_ids_file_path: Union[str, os.PathLike[Any]] = "", + config: Optional[Coqpit] = None, ): super().__init__(id_file_path=language_ids_file_path) @@ -59,7 +59,7 @@ def parse_language_ids_from_config(c: Coqpit) -> Dict: languages.add(dataset["language"]) else: raise ValueError(f"Dataset {dataset['name']} has no language specified.") - return {name: i for i, name in enumerate(sorted(list(languages)))} + return {name: i for i, name in enumerate(sorted(languages))} def set_language_ids_from_config(self, c: Coqpit) -> None: """Set language IDs from config samples. @@ -76,7 +76,7 @@ def parse_ids_from_data(items: List, parse_key: str) -> Any: def set_ids_from_data(self, items: List, parse_key: str) -> Any: raise NotImplementedError - def save_ids_to_file(self, file_path: str) -> None: + def save_ids_to_file(self, file_path: Union[str, os.PathLike[Any]]) -> None: """Save language IDs to a json file. Args: @@ -85,18 +85,18 @@ def save_ids_to_file(self, file_path: str) -> None: self._save_json(file_path, self.name_to_id) @staticmethod - def init_from_config(config: Coqpit) -> "LanguageManager": + def init_from_config(config: Coqpit) -> Optional["LanguageManager"]: """Initialize the language manager from a Coqpit config. Args: config (Coqpit): Coqpit config. """ - language_manager = None if check_config_and_model_args(config, "use_language_embedding", True): if config.get("language_ids_file", None): - language_manager = LanguageManager(language_ids_file_path=config.language_ids_file) - language_manager = LanguageManager(config=config) - return language_manager + return LanguageManager(language_ids_file_path=config.language_ids_file) + # Fall back to parse language IDs from the config + return LanguageManager(config=config) + return None def _set_file_path(path): diff --git a/TTS/tts/utils/managers.py b/TTS/tts/utils/managers.py index 1f94c5332d..3a715dd75d 100644 --- a/TTS/tts/utils/managers.py +++ b/TTS/tts/utils/managers.py @@ -1,4 +1,5 @@ import json +import os import random from typing import Any, Dict, List, Tuple, Union @@ -9,20 +10,23 @@ from TTS.config import load_config from TTS.encoder.utils.generic_utils import setup_encoder_model from TTS.utils.audio import AudioProcessor +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 -def load_file(path: str): +def load_file(path: Union[str, os.PathLike[Any]]): + path = str(path) if path.endswith(".json"): with fsspec.open(path, "r") as f: return json.load(f) elif path.endswith(".pth"): with fsspec.open(path, "rb") as f: - return torch.load(f, map_location="cpu") + return torch.load(f, map_location="cpu", weights_only=is_pytorch_at_least_2_4()) else: raise ValueError("Unsupported file type") -def save_file(obj: Any, path: str): +def save_file(obj: Any, path: Union[str, os.PathLike[Any]]): + path = str(path) if path.endswith(".json"): with fsspec.open(path, "w") as f: json.dump(obj, f, indent=4) @@ -38,20 +42,20 @@ class BaseIDManager: It defines common `ID` manager specific functions. """ - def __init__(self, id_file_path: str = ""): + def __init__(self, id_file_path: Union[str, os.PathLike[Any]] = ""): self.name_to_id = {} if id_file_path: self.load_ids_from_file(id_file_path) @staticmethod - def _load_json(json_file_path: str) -> Dict: - with fsspec.open(json_file_path, "r") as f: + def _load_json(json_file_path: Union[str, os.PathLike[Any]]) -> Dict: + with fsspec.open(str(json_file_path), "r") as f: return json.load(f) @staticmethod - def _save_json(json_file_path: str, data: dict) -> None: - with fsspec.open(json_file_path, "w") as f: + def _save_json(json_file_path: Union[str, os.PathLike[Any]], data: dict) -> None: + with fsspec.open(str(json_file_path), "w") as f: json.dump(data, f, indent=4) def set_ids_from_data(self, items: List, parse_key: str) -> None: @@ -62,7 +66,7 @@ def set_ids_from_data(self, items: List, parse_key: str) -> None: """ self.name_to_id = self.parse_ids_from_data(items, parse_key=parse_key) - def load_ids_from_file(self, file_path: str) -> None: + def load_ids_from_file(self, file_path: Union[str, os.PathLike[Any]]) -> None: """Set IDs from a file. Args: @@ -70,7 +74,7 @@ def load_ids_from_file(self, file_path: str) -> None: """ self.name_to_id = load_file(file_path) - def save_ids_to_file(self, file_path: str) -> None: + def save_ids_to_file(self, file_path: Union[str, os.PathLike[Any]]) -> None: """Save IDs to a json file. Args: @@ -129,10 +133,10 @@ class EmbeddingManager(BaseIDManager): def __init__( self, - embedding_file_path: Union[str, List[str]] = "", - id_file_path: str = "", - encoder_model_path: str = "", - encoder_config_path: str = "", + embedding_file_path: Union[Union[str, os.PathLike[Any]], list[Union[str, os.PathLike[Any]]]] = "", + id_file_path: Union[str, os.PathLike[Any]] = "", + encoder_model_path: Union[str, os.PathLike[Any]] = "", + encoder_config_path: Union[str, os.PathLike[Any]] = "", use_cuda: bool = False, ): super().__init__(id_file_path=id_file_path) @@ -175,7 +179,7 @@ def embedding_names(self): """Get embedding names.""" return list(self.embeddings_by_names.keys()) - def save_embeddings_to_file(self, file_path: str) -> None: + def save_embeddings_to_file(self, file_path: Union[str, os.PathLike[Any]]) -> None: """Save embeddings to a json file. Args: @@ -184,7 +188,7 @@ def save_embeddings_to_file(self, file_path: str) -> None: save_file(self.embeddings, file_path) @staticmethod - def read_embeddings_from_file(file_path: str): + def read_embeddings_from_file(file_path: Union[str, os.PathLike[Any]]): """Load embeddings from a json file. Args: @@ -193,7 +197,7 @@ def read_embeddings_from_file(file_path: str): embeddings = load_file(file_path) speakers = sorted({x["name"] for x in embeddings.values()}) name_to_id = {name: i for i, name in enumerate(speakers)} - clip_ids = list(set(sorted(clip_name for clip_name in embeddings.keys()))) + clip_ids = list(set(clip_name for clip_name in embeddings.keys())) # cache embeddings_by_names for fast inference using a bigger speakers.json embeddings_by_names = {} for x in embeddings.values(): @@ -203,7 +207,7 @@ def read_embeddings_from_file(file_path: str): embeddings_by_names[x["name"]].append(x["embedding"]) return name_to_id, clip_ids, embeddings, embeddings_by_names - def load_embeddings_from_file(self, file_path: str) -> None: + def load_embeddings_from_file(self, file_path: Union[str, os.PathLike[Any]]) -> None: """Load embeddings from a json file. Args: @@ -213,7 +217,7 @@ def load_embeddings_from_file(self, file_path: str) -> None: file_path ) - def load_embeddings_from_list_of_files(self, file_paths: List[str]) -> None: + def load_embeddings_from_list_of_files(self, file_paths: list[Union[str, os.PathLike[Any]]]) -> None: """Load embeddings from a list of json files and don't allow duplicate keys. Args: @@ -312,7 +316,9 @@ def get_random_embedding(self) -> Any: def get_clips(self) -> List: return sorted(self.embeddings.keys()) - def init_encoder(self, model_path: str, config_path: str, use_cuda=False) -> None: + def init_encoder( + self, model_path: Union[str, os.PathLike[Any]], config_path: Union[str, os.PathLike[Any]], use_cuda=False + ) -> None: """Initialize a speaker encoder model. Args: @@ -324,11 +330,13 @@ def init_encoder(self, model_path: str, config_path: str, use_cuda=False) -> Non self.encoder_config = load_config(config_path) self.encoder = setup_encoder_model(self.encoder_config) self.encoder_criterion = self.encoder.load_checkpoint( - self.encoder_config, model_path, eval=True, use_cuda=use_cuda, cache=True + self.encoder_config, str(model_path), eval=True, use_cuda=use_cuda, cache=True ) self.encoder_ap = AudioProcessor(**self.encoder_config.audio) - def compute_embedding_from_clip(self, wav_file: Union[str, List[str]]) -> list: + def compute_embedding_from_clip( + self, wav_file: Union[Union[str, os.PathLike[Any]], List[Union[str, os.PathLike[Any]]]] + ) -> list: """Compute a embedding from a given audio file. Args: diff --git a/TTS/tts/utils/monotonic_align/core.pyx b/TTS/tts/utils/monotonic_align/core.pyx deleted file mode 100644 index 091fcc3a50..0000000000 --- a/TTS/tts/utils/monotonic_align/core.pyx +++ /dev/null @@ -1,47 +0,0 @@ -import numpy as np - -cimport cython -cimport numpy as np - -from cython.parallel import prange - - -@cython.boundscheck(False) -@cython.wraparound(False) -cdef void maximum_path_each(int[:,::1] path, float[:,::1] value, int t_x, int t_y, float max_neg_val) nogil: - cdef int x - cdef int y - cdef float v_prev - cdef float v_cur - cdef float tmp - cdef int index = t_x - 1 - - for y in range(t_y): - for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)): - if x == y: - v_cur = max_neg_val - else: - v_cur = value[x, y-1] - if x == 0: - if y == 0: - v_prev = 0. - else: - v_prev = max_neg_val - else: - v_prev = value[x-1, y-1] - value[x, y] = max(v_cur, v_prev) + value[x, y] - - for y in range(t_y - 1, -1, -1): - path[index, y] = 1 - if index != 0 and (index == y or value[index, y-1] < value[index-1, y-1]): - index = index - 1 - - -@cython.boundscheck(False) -@cython.wraparound(False) -cpdef void maximum_path_c(int[:,:,::1] paths, float[:,:,::1] values, int[::1] t_xs, int[::1] t_ys, float max_neg_val=-1e9) nogil: - cdef int b = values.shape[0] - - cdef int i - for i in prange(b, nogil=True): - maximum_path_each(paths[i], values[i], t_xs[i], t_ys[i], max_neg_val) diff --git a/TTS/tts/utils/monotonic_align/setup.py b/TTS/tts/utils/monotonic_align/setup.py deleted file mode 100644 index f22bc6a35a..0000000000 --- a/TTS/tts/utils/monotonic_align/setup.py +++ /dev/null @@ -1,7 +0,0 @@ -# from distutils.core import setup -# from Cython.Build import cythonize -# import numpy - -# setup(name='monotonic_align', -# ext_modules=cythonize("core.pyx"), -# include_dirs=[numpy.get_include()]) diff --git a/TTS/tts/utils/speakers.py b/TTS/tts/utils/speakers.py index e49695268d..89c56583f5 100644 --- a/TTS/tts/utils/speakers.py +++ b/TTS/tts/utils/speakers.py @@ -1,6 +1,7 @@ import json +import logging import os -from typing import Any, Dict, List, Union +from typing import Any, Dict, List, Optional, Union import fsspec import numpy as np @@ -10,6 +11,8 @@ from TTS.config import get_from_config_or_model_args_with_default from TTS.tts.utils.managers import EmbeddingManager +logger = logging.getLogger(__name__) + class SpeakerManager(EmbeddingManager): """Manage the speakers for multi-speaker 🐸TTS models. Load a datafile and parse the information @@ -53,11 +56,11 @@ class SpeakerManager(EmbeddingManager): def __init__( self, - data_items: List[List[Any]] = None, + data_items: Optional[list[list[Any]]] = None, d_vectors_file_path: str = "", - speaker_id_file_path: str = "", - encoder_model_path: str = "", - encoder_config_path: str = "", + speaker_id_file_path: Union[str, os.PathLike[Any]] = "", + encoder_model_path: Union[str, os.PathLike[Any]] = "", + encoder_config_path: Union[str, os.PathLike[Any]] = "", use_cuda: bool = False, ): super().__init__( @@ -170,7 +173,9 @@ def get_speaker_manager(c: Coqpit, data: List = None, restore_path: str = None, if c.use_d_vector_file: # restore speaker manager with the embedding file if not os.path.exists(speakers_file): - print("WARNING: speakers.json was not found in restore_path, trying to use CONFIG.d_vector_file") + logger.warning( + "speakers.json was not found in %s, trying to use CONFIG.d_vector_file", restore_path + ) if not os.path.exists(c.d_vector_file): raise RuntimeError( "You must copy the file speakers.json to restore_path, or set a valid file in CONFIG.d_vector_file" @@ -193,16 +198,16 @@ def get_speaker_manager(c: Coqpit, data: List = None, restore_path: str = None, speaker_manager.load_ids_from_file(c.speakers_file) if speaker_manager.num_speakers > 0: - print( - " > Speaker manager is loaded with {} speakers: {}".format( - speaker_manager.num_speakers, ", ".join(speaker_manager.name_to_id) - ) + logger.info( + "Speaker manager is loaded with %d speakers: %s", + speaker_manager.num_speakers, + ", ".join(speaker_manager.name_to_id), ) # save file if path is defined if out_path: out_file_path = os.path.join(out_path, "speakers.json") - print(f" > Saving `speakers.json` to {out_file_path}.") + logger.info("Saving `speakers.json` to %s", out_file_path) if c.use_d_vector_file and c.d_vector_file: speaker_manager.save_embeddings_to_file(out_file_path) else: diff --git a/TTS/tts/utils/ssim.py b/TTS/tts/utils/ssim.py index 4bc3befc5b..eddf05db3f 100644 --- a/TTS/tts/utils/ssim.py +++ b/TTS/tts/utils/ssim.py @@ -207,6 +207,7 @@ class SSIMLoss(_Loss): https://ece.uwaterloo.ca/~z70wang/publications/ssim.pdf, DOI:`10.1109/TIP.2003.819861` """ + __constants__ = ["kernel_size", "k1", "k2", "sigma", "kernel", "reduction"] def __init__( diff --git a/TTS/tts/utils/synthesis.py b/TTS/tts/utils/synthesis.py index 797151c254..5dc4cc569f 100644 --- a/TTS/tts/utils/synthesis.py +++ b/TTS/tts/utils/synthesis.py @@ -1,17 +1,16 @@ -from typing import Dict +from typing import Dict, Optional, Union import numpy as np import torch from torch import nn -def numpy_to_torch(np_array, dtype, cuda=False, device="cpu"): - if cuda: - device = "cuda" +def numpy_to_torch( + np_array: np.ndarray, dtype: torch.dtype, device: Union[str, torch.device] = "cpu" +) -> Optional[torch.Tensor]: if np_array is None: return None - tensor = torch.as_tensor(np_array, dtype=dtype, device=device) - return tensor + return torch.as_tensor(np_array, dtype=dtype, device=device) def compute_style_mel(style_wav, ap, cuda=False, device="cpu"): @@ -76,18 +75,14 @@ def inv_spectrogram(postnet_output, ap, CONFIG): return wav -def id_to_torch(aux_id, cuda=False, device="cpu"): - if cuda: - device = "cuda" +def id_to_torch(aux_id, device: Union[str, torch.device] = "cpu") -> Optional[torch.Tensor]: if aux_id is not None: aux_id = np.asarray(aux_id) aux_id = torch.from_numpy(aux_id).to(device) return aux_id -def embedding_to_torch(d_vector, cuda=False, device="cpu"): - if cuda: - device = "cuda" +def embedding_to_torch(d_vector, device: Union[str, torch.device] = "cpu") -> Optional[torch.Tensor]: if d_vector is not None: d_vector = np.asarray(d_vector) d_vector = torch.from_numpy(d_vector).type(torch.FloatTensor) diff --git a/TTS/tts/utils/text/bangla/phonemizer.py b/TTS/tts/utils/text/bangla/phonemizer.py index e15830fe8a..cddcb00fd5 100644 --- a/TTS/tts/utils/text/bangla/phonemizer.py +++ b/TTS/tts/utils/text/bangla/phonemizer.py @@ -1,8 +1,11 @@ import re -import bangla -from bnnumerizer import numerize -from bnunicodenormalizer import Normalizer +try: + import bangla + from bnnumerizer import numerize + from bnunicodenormalizer import Normalizer +except ImportError as e: + raise ImportError("Bangla requires: bangla, bnnumerizer, bnunicodenormalizer") from e # initialize bnorm = Normalizer() diff --git a/TTS/tts/utils/text/characters.py b/TTS/tts/utils/text/characters.py index 8fa45ed84b..4bf9bf6bd5 100644 --- a/TTS/tts/utils/text/characters.py +++ b/TTS/tts/utils/text/characters.py @@ -1,8 +1,11 @@ +import logging from dataclasses import replace from typing import Dict from TTS.tts.configs.shared_configs import CharactersConfig +logger = logging.getLogger(__name__) + def parse_symbols(): return { @@ -31,8 +34,8 @@ def parse_symbols(): _pulmonic_consonants = "pbtdʈɖcɟkÉĄqÉĸĘ”É´Å‹É˛ÉŗnÉąmʙrʀⱱɾÉŊɸβfvθðszĘƒĘ’Ę‚ĘÃ§ĘxÉŖĪ‡ĘÄ§Ę•hÉĻÉŦɮʋɹÉģjÉ°lɭʎʟ" _suprasegmentals = "ˈˌːˑ" _other_symbols = "ʍwÉĨʜĘĸʡɕʑÉēɧʲ" -_diacrilics = "ɚ˞ÉĢ" -_phonemes = _vowels + _non_pulmonic_consonants + _pulmonic_consonants + _suprasegmentals + _other_symbols + _diacrilics +_diacritics = "ĖƒÉšËžÉĢ" +_phonemes = _vowels + _non_pulmonic_consonants + _pulmonic_consonants + _suprasegmentals + _other_symbols + _diacritics class BaseVocabulary: @@ -87,9 +90,7 @@ def vocab(self, vocab): if vocab is not None: self._vocab = vocab self._char_to_id = {char: idx for idx, char in enumerate(self._vocab)} - self._id_to_char = { - idx: char for idx, char in enumerate(self._vocab) # pylint: disable=unnecessary-comprehension - } + self._id_to_char = dict(enumerate(self._vocab)) @staticmethod def init_from_config(config, **kwargs): @@ -269,9 +270,7 @@ def vocab(self): def vocab(self, vocab): self._vocab = vocab self._char_to_id = {char: idx for idx, char in enumerate(self.vocab)} - self._id_to_char = { - idx: char for idx, char in enumerate(self.vocab) # pylint: disable=unnecessary-comprehension - } + self._id_to_char = dict(enumerate(self.vocab)) @property def num_chars(self): @@ -309,14 +308,14 @@ def print_log(self, level: int = 0): Prints the vocabulary in a nice format. """ indent = "\t" * level - print(f"{indent}| > Characters: {self._characters}") - print(f"{indent}| > Punctuations: {self._punctuations}") - print(f"{indent}| > Pad: {self._pad}") - print(f"{indent}| > EOS: {self._eos}") - print(f"{indent}| > BOS: {self._bos}") - print(f"{indent}| > Blank: {self._blank}") - print(f"{indent}| > Vocab: {self.vocab}") - print(f"{indent}| > Num chars: {self.num_chars}") + logger.info("%s| Characters: %s", indent, self._characters) + logger.info("%s| Punctuations: %s", indent, self._punctuations) + logger.info("%s| Pad: %s", indent, self._pad) + logger.info("%s| EOS: %s", indent, self._eos) + logger.info("%s| BOS: %s", indent, self._bos) + logger.info("%s| Blank: %s", indent, self._blank) + logger.info("%s| Vocab: %s", indent, self.vocab) + logger.info("%s| Num chars: %d", indent, self.num_chars) @staticmethod def init_from_config(config: "Coqpit"): # pylint: disable=unused-argument diff --git a/TTS/tts/utils/text/chinese_mandarin/phonemizer.py b/TTS/tts/utils/text/chinese_mandarin/phonemizer.py index 727c881e10..e9d62e9d06 100644 --- a/TTS/tts/utils/text/chinese_mandarin/phonemizer.py +++ b/TTS/tts/utils/text/chinese_mandarin/phonemizer.py @@ -1,7 +1,10 @@ from typing import List -import jieba -import pypinyin +try: + import jieba + import pypinyin +except ImportError as e: + raise ImportError("Chinese requires: jieba, pypinyin") from e from .pinyinToPhonemes import PINYIN_DICT diff --git a/TTS/tts/utils/text/chinese_mandarin/pinyinToPhonemes.py b/TTS/tts/utils/text/chinese_mandarin/pinyinToPhonemes.py index 4e25c3a4c9..89dd654ab1 100644 --- a/TTS/tts/utils/text/chinese_mandarin/pinyinToPhonemes.py +++ b/TTS/tts/utils/text/chinese_mandarin/pinyinToPhonemes.py @@ -94,25 +94,25 @@ "fo": ["fo"], "fou": ["fou"], "fu": ["fu"], - "ga": ["ga"], - "gai": ["gai"], - "gan": ["gan"], - "gang": ["gɑŋ"], - "gao": ["gaʌ"], - "ge": ["gø"], - "gei": ["gei"], - "gen": ["gœn"], - "geng": ["gÉĩŋ"], - "gong": ["goŋ"], - "gou": ["gou"], - "gu": ["gu"], - "gua": ["gua"], - "guai": ["guai"], - "guan": ["guan"], - "guang": ["guɑŋ"], - "gui": ["guei"], - "gun": ["gun"], - "guo": ["guo"], + "ga": ["ÉĄa"], + "gai": ["ÉĄai"], + "gan": ["ÉĄan"], + "gang": ["ÉĄÉ‘Å‹"], + "gao": ["ÉĄaʌ"], + "ge": ["ÉĄÃ¸"], + "gei": ["ÉĄei"], + "gen": ["ÉĄÅ“n"], + "geng": ["ÉĄÉĩŋ"], + "gong": ["ÉĄoŋ"], + "gou": ["ÉĄou"], + "gu": ["ÉĄu"], + "gua": ["ÉĄua"], + "guai": ["ÉĄuai"], + "guan": ["ÉĄuan"], + "guang": ["ÉĄuɑŋ"], + "gui": ["ÉĄuei"], + "gun": ["ÉĄun"], + "guo": ["ÉĄuo"], "ha": ["xa"], "hai": ["xai"], "han": ["xan"], diff --git a/TTS/tts/utils/text/cleaners.py b/TTS/tts/utils/text/cleaners.py index 74d3910b51..f496b9f0dd 100644 --- a/TTS/tts/utils/text/cleaners.py +++ b/TTS/tts/utils/text/cleaners.py @@ -1,7 +1,8 @@ """Set of default text cleaners""" -# TODO: pick the cleaner for languages dynamically import re +from typing import Optional +from unicodedata import normalize from anyascii import anyascii @@ -16,35 +17,38 @@ _whitespace_re = re.compile(r"\s+") -def expand_abbreviations(text, lang="en"): +def expand_abbreviations(text: str, lang: str = "en") -> str: if lang == "en": _abbreviations = abbreviations_en elif lang == "fr": _abbreviations = abbreviations_fr + else: + msg = f"Language {lang} not supported in expand_abbreviations" + raise ValueError(msg) for regex, replacement in _abbreviations: text = re.sub(regex, replacement, text) return text -def lowercase(text): +def lowercase(text: str) -> str: return text.lower() -def collapse_whitespace(text): +def collapse_whitespace(text: str) -> str: return re.sub(_whitespace_re, " ", text).strip() -def convert_to_ascii(text): +def convert_to_ascii(text: str) -> str: return anyascii(text) -def remove_aux_symbols(text): +def remove_aux_symbols(text: str) -> str: text = re.sub(r"[\<\>\(\)\[\]\"]+", "", text) return text -def replace_symbols(text, lang="en"): - """Replace symbols based on the lenguage tag. +def replace_symbols(text: str, lang: Optional[str] = "en") -> str: + """Replace symbols based on the language tag. Args: text: @@ -76,39 +80,44 @@ def replace_symbols(text, lang="en"): return text -def basic_cleaners(text): +def basic_cleaners(text: str) -> str: """Basic pipeline that lowercases and collapses whitespace without transliteration.""" + text = normalize_unicode(text) text = lowercase(text) text = collapse_whitespace(text) return text -def transliteration_cleaners(text): +def transliteration_cleaners(text: str) -> str: """Pipeline for non-English text that transliterates to ASCII.""" + text = normalize_unicode(text) # text = convert_to_ascii(text) text = lowercase(text) text = collapse_whitespace(text) return text -def basic_german_cleaners(text): +def basic_german_cleaners(text: str) -> str: """Pipeline for German text""" + text = normalize_unicode(text) text = lowercase(text) text = collapse_whitespace(text) return text # TODO: elaborate it -def basic_turkish_cleaners(text): +def basic_turkish_cleaners(text: str) -> str: """Pipeline for Turkish text""" + text = normalize_unicode(text) text = text.replace("I", "Äą") text = lowercase(text) text = collapse_whitespace(text) return text -def english_cleaners(text): +def english_cleaners(text: str) -> str: """Pipeline for English text, including number and abbreviation expansion.""" + text = normalize_unicode(text) # text = convert_to_ascii(text) text = lowercase(text) text = expand_time_english(text) @@ -120,8 +129,13 @@ def english_cleaners(text): return text -def phoneme_cleaners(text): - """Pipeline for phonemes mode, including number and abbreviation expansion.""" +def phoneme_cleaners(text: str) -> str: + """Pipeline for phonemes mode, including number and abbreviation expansion. + + NB: This cleaner converts numbers into English words, for other languages + use multilingual_phoneme_cleaners(). + """ + text = normalize_unicode(text) text = en_normalize_numbers(text) text = expand_abbreviations(text) text = replace_symbols(text) @@ -130,8 +144,18 @@ def phoneme_cleaners(text): return text -def french_cleaners(text): +def multilingual_phoneme_cleaners(text: str) -> str: + """Pipeline for phonemes mode, including number and abbreviation expansion.""" + text = normalize_unicode(text) + text = replace_symbols(text, lang=None) + text = remove_aux_symbols(text) + text = collapse_whitespace(text) + return text + + +def french_cleaners(text: str) -> str: """Pipeline for French text. There is no need to expand numbers, phonemizer already does that""" + text = normalize_unicode(text) text = expand_abbreviations(text, lang="fr") text = lowercase(text) text = replace_symbols(text, lang="fr") @@ -140,9 +164,10 @@ def french_cleaners(text): return text -def portuguese_cleaners(text): +def portuguese_cleaners(text: str) -> str: """Basic pipeline for Portuguese text. There is no need to expand abbreviation and numbers, phonemizer already does that""" + text = normalize_unicode(text) text = lowercase(text) text = replace_symbols(text, lang="pt") text = remove_aux_symbols(text) @@ -152,12 +177,14 @@ def portuguese_cleaners(text): def chinese_mandarin_cleaners(text: str) -> str: """Basic pipeline for chinese""" + text = normalize_unicode(text) text = replace_numbers_to_characters_in_text(text) return text -def multilingual_cleaners(text): +def multilingual_cleaners(text: str) -> str: """Pipeline for multilingual text""" + text = normalize_unicode(text) text = lowercase(text) text = replace_symbols(text, lang=None) text = remove_aux_symbols(text) @@ -165,7 +192,13 @@ def multilingual_cleaners(text): return text -def no_cleaners(text): +def no_cleaners(text: str) -> str: # remove newline characters text = text.replace("\n", "") return text + + +def normalize_unicode(text: str) -> str: + """Normalize Unicode characters.""" + text = normalize("NFC", text) + return text diff --git a/TTS/tts/utils/text/japanese/phonemizer.py b/TTS/tts/utils/text/japanese/phonemizer.py index c3111067e1..30072ae501 100644 --- a/TTS/tts/utils/text/japanese/phonemizer.py +++ b/TTS/tts/utils/text/japanese/phonemizer.py @@ -350,8 +350,8 @@ def hira2kata(text: str) -> str: return text.replace("う゛", "ヴ") -_SYMBOL_TOKENS = set(list("ãƒģ、。īŧŸīŧ")) -_NO_YOMI_TOKENS = set(list("「」『』―īŧˆīŧ‰īŧģīŧŊ[]　â€Ļ")) +_SYMBOL_TOKENS = set("ãƒģ、。īŧŸīŧ") +_NO_YOMI_TOKENS = set("「」『』―īŧˆīŧ‰īŧģīŧŊ[]　â€Ļ") _TAGGER = MeCab.Tagger() diff --git a/TTS/tts/utils/text/korean/phonemizer.py b/TTS/tts/utils/text/korean/phonemizer.py index 2c69217c40..dde039b0f5 100644 --- a/TTS/tts/utils/text/korean/phonemizer.py +++ b/TTS/tts/utils/text/korean/phonemizer.py @@ -1,4 +1,7 @@ -from jamo import hangul_to_jamo +try: + from jamo import hangul_to_jamo +except ImportError as e: + raise ImportError("Korean requires: g2pkk, jamo") from e from TTS.tts.utils.text.korean.korean import normalize diff --git a/TTS/tts/utils/text/phonemizers/__init__.py b/TTS/tts/utils/text/phonemizers/__init__.py index f9a0340c55..fdf62bab3d 100644 --- a/TTS/tts/utils/text/phonemizers/__init__.py +++ b/TTS/tts/utils/text/phonemizers/__init__.py @@ -1,18 +1,29 @@ -from TTS.tts.utils.text.phonemizers.bangla_phonemizer import BN_Phonemizer from TTS.tts.utils.text.phonemizers.base import BasePhonemizer from TTS.tts.utils.text.phonemizers.belarusian_phonemizer import BEL_Phonemizer from TTS.tts.utils.text.phonemizers.espeak_wrapper import ESpeak from TTS.tts.utils.text.phonemizers.gruut_wrapper import Gruut -from TTS.tts.utils.text.phonemizers.ko_kr_phonemizer import KO_KR_Phonemizer -from TTS.tts.utils.text.phonemizers.zh_cn_phonemizer import ZH_CN_Phonemizer + +try: + from TTS.tts.utils.text.phonemizers.bangla_phonemizer import BN_Phonemizer +except ImportError: + BN_Phonemizer = None try: from TTS.tts.utils.text.phonemizers.ja_jp_phonemizer import JA_JP_Phonemizer except ImportError: JA_JP_Phonemizer = None - pass -PHONEMIZERS = {b.name(): b for b in (ESpeak, Gruut, KO_KR_Phonemizer, BN_Phonemizer)} +try: + from TTS.tts.utils.text.phonemizers.ko_kr_phonemizer import KO_KR_Phonemizer +except ImportError: + KO_KR_Phonemizer = None + +try: + from TTS.tts.utils.text.phonemizers.zh_cn_phonemizer import ZH_CN_Phonemizer +except ImportError: + ZH_CN_Phonemizer = None + +PHONEMIZERS = {b.name(): b for b in (ESpeak, Gruut)} ESPEAK_LANGS = list(ESpeak.supported_languages().keys()) @@ -33,17 +44,21 @@ # Force default for some languages DEF_LANG_TO_PHONEMIZER["en"] = DEF_LANG_TO_PHONEMIZER["en-us"] -DEF_LANG_TO_PHONEMIZER["zh-cn"] = ZH_CN_Phonemizer.name() -DEF_LANG_TO_PHONEMIZER["ko-kr"] = KO_KR_Phonemizer.name() -DEF_LANG_TO_PHONEMIZER["bn"] = BN_Phonemizer.name() DEF_LANG_TO_PHONEMIZER["be"] = BEL_Phonemizer.name() -# JA phonemizer has deal breaking dependencies like MeCab for some systems. -# So we only have it when we have it. +if BN_Phonemizer is not None: + PHONEMIZERS[BN_Phonemizer.name()] = BN_Phonemizer + DEF_LANG_TO_PHONEMIZER["bn"] = BN_Phonemizer.name() if JA_JP_Phonemizer is not None: PHONEMIZERS[JA_JP_Phonemizer.name()] = JA_JP_Phonemizer DEF_LANG_TO_PHONEMIZER["ja-jp"] = JA_JP_Phonemizer.name() +if KO_KR_Phonemizer is not None: + PHONEMIZERS[KO_KR_Phonemizer.name()] = KO_KR_Phonemizer + DEF_LANG_TO_PHONEMIZER["ko-kr"] = KO_KR_Phonemizer.name() +if ZH_CN_Phonemizer is not None: + PHONEMIZERS[ZH_CN_Phonemizer.name()] = ZH_CN_Phonemizer + DEF_LANG_TO_PHONEMIZER["zh-cn"] = ZH_CN_Phonemizer.name() def get_phonemizer_by_name(name: str, **kwargs) -> BasePhonemizer: @@ -61,14 +76,20 @@ def get_phonemizer_by_name(name: str, **kwargs) -> BasePhonemizer: if name == "gruut": return Gruut(**kwargs) if name == "zh_cn_phonemizer": + if ZH_CN_Phonemizer is None: + raise ValueError("You need to install ZH phonemizer dependencies. Try `pip install coqui-tts[zh]`.") return ZH_CN_Phonemizer(**kwargs) if name == "ja_jp_phonemizer": if JA_JP_Phonemizer is None: - raise ValueError(" ❗ You need to install JA phonemizer dependencies. Try `pip install TTS[ja]`.") + raise ValueError("You need to install JA phonemizer dependencies. Try `pip install coqui-tts[ja]`.") return JA_JP_Phonemizer(**kwargs) if name == "ko_kr_phonemizer": + if KO_KR_Phonemizer is None: + raise ValueError("You need to install KO phonemizer dependencies. Try `pip install coqui-tts[ko]`.") return KO_KR_Phonemizer(**kwargs) if name == "bn_phonemizer": + if BN_Phonemizer is None: + raise ValueError("You need to install BN phonemizer dependencies. Try `pip install coqui-tts[bn]`.") return BN_Phonemizer(**kwargs) if name == "be_phonemizer": return BEL_Phonemizer(**kwargs) diff --git a/TTS/tts/utils/text/phonemizers/base.py b/TTS/tts/utils/text/phonemizers/base.py index 4fc7987415..5e701df458 100644 --- a/TTS/tts/utils/text/phonemizers/base.py +++ b/TTS/tts/utils/text/phonemizers/base.py @@ -1,8 +1,11 @@ import abc +import logging from typing import List, Tuple from TTS.tts.utils.text.punctuation import Punctuation +logger = logging.getLogger(__name__) + class BasePhonemizer(abc.ABC): """Base phonemizer class @@ -136,5 +139,5 @@ def phonemize(self, text: str, separator="|", language: str = None) -> str: # p def print_logs(self, level: int = 0): indent = "\t" * level - print(f"{indent}| > phoneme language: {self.language}") - print(f"{indent}| > phoneme backend: {self.name()}") + logger.info("%s| phoneme language: %s", indent, self.language) + logger.info("%s| phoneme backend: %s", indent, self.name()) diff --git a/TTS/tts/utils/text/phonemizers/espeak_wrapper.py b/TTS/tts/utils/text/phonemizers/espeak_wrapper.py index 328e52f369..a15df716e7 100644 --- a/TTS/tts/utils/text/phonemizers/espeak_wrapper.py +++ b/TTS/tts/utils/text/phonemizers/espeak_wrapper.py @@ -1,15 +1,21 @@ +"""Wrapper to call the espeak/espeak-ng phonemizer.""" + import logging import re import subprocess -from typing import Dict, List +import tempfile +from pathlib import Path +from typing import Optional from packaging.version import Version from TTS.tts.utils.text.phonemizers.base import BasePhonemizer from TTS.tts.utils.text.punctuation import Punctuation +logger = logging.getLogger(__name__) + -def is_tool(name): +def _is_tool(name) -> bool: from shutil import which return which(name) is not None @@ -20,23 +26,25 @@ def is_tool(name): espeak_version_pattern = re.compile(r"text-to-speech:\s(?P\d+\.\d+(\.\d+)?)") -def get_espeak_version(): +def get_espeak_version() -> str: + """Return version of the `espeak` binary.""" output = subprocess.getoutput("espeak --version") match = espeak_version_pattern.search(output) return match.group("version") -def get_espeakng_version(): +def get_espeakng_version() -> str: + """Return version of the `espeak-ng` binary.""" output = subprocess.getoutput("espeak-ng --version") return output.split()[3] # priority: espeakng > espeak -if is_tool("espeak-ng"): +if _is_tool("espeak-ng"): _DEF_ESPEAK_LIB = "espeak-ng" _DEF_ESPEAK_VER = get_espeakng_version() -elif is_tool("espeak"): +elif _is_tool("espeak"): _DEF_ESPEAK_LIB = "espeak" _DEF_ESPEAK_VER = get_espeak_version() else: @@ -44,7 +52,7 @@ def get_espeakng_version(): _DEF_ESPEAK_VER = None -def _espeak_exe(espeak_lib: str, args: List, sync=False) -> List[str]: +def _espeak_exe(espeak_lib: str, args: list) -> list[str]: """Run espeak with the given arguments.""" cmd = [ espeak_lib, @@ -53,35 +61,22 @@ def _espeak_exe(espeak_lib: str, args: List, sync=False) -> List[str]: "1", # UTF8 text encoding ] cmd.extend(args) - logging.debug("espeakng: executing %s", repr(cmd)) - - with subprocess.Popen( - cmd, - stdout=subprocess.PIPE, - stderr=subprocess.STDOUT, - ) as p: - res = iter(p.stdout.readline, b"") - if not sync: - p.stdout.close() - if p.stderr: - p.stderr.close() - if p.stdin: - p.stdin.close() - return res - res2 = [] - for line in res: - res2.append(line) - p.stdout.close() - if p.stderr: - p.stderr.close() - if p.stdin: - p.stdin.close() - p.wait() - return res2 + logger.debug("Executing: %s", repr(cmd)) + + p = subprocess.run(cmd, capture_output=True, encoding="utf8", check=True) + for line in p.stderr.strip().split("\n"): + if line.strip() != "": + logger.warning("%s: %s", espeak_lib, line.strip()) + res = [] + for line in p.stdout.strip().split("\n"): + if line.strip() != "": + logger.debug("%s: %s", espeak_lib, line.strip()) + res.append(line.strip()) + return res class ESpeak(BasePhonemizer): - """ESpeak wrapper calling `espeak` or `espeak-ng` from the command-line the perform G2P + """Wrapper calling `espeak` or `espeak-ng` from the command-line to perform G2P. Args: language (str): @@ -106,13 +101,17 @@ class ESpeak(BasePhonemizer): """ - _ESPEAK_LIB = _DEF_ESPEAK_LIB - _ESPEAK_VER = _DEF_ESPEAK_VER - - def __init__(self, language: str, backend=None, punctuations=Punctuation.default_puncs(), keep_puncs=True): - if self._ESPEAK_LIB is None: - raise Exception(" [!] No espeak backend found. Install espeak-ng or espeak to your system.") - self.backend = self._ESPEAK_LIB + def __init__( + self, + language: str, + backend: Optional[str] = None, + punctuations: str = Punctuation.default_puncs(), + keep_puncs: bool = True, + ): + if _DEF_ESPEAK_LIB is None: + msg = "[!] No espeak backend found. Install espeak-ng or espeak to your system." + raise FileNotFoundError(msg) + self.backend = _DEF_ESPEAK_LIB # band-aid for backwards compatibility if language == "en": @@ -125,35 +124,37 @@ def __init__(self, language: str, backend=None, punctuations=Punctuation.default self.backend = backend @property - def backend(self): + def backend(self) -> str: return self._ESPEAK_LIB @property - def backend_version(self): + def backend_version(self) -> str: return self._ESPEAK_VER @backend.setter - def backend(self, backend): + def backend(self, backend: str) -> None: if backend not in ["espeak", "espeak-ng"]: - raise Exception("Unknown backend: %s" % backend) + msg = f"Unknown backend: {backend}" + raise ValueError(msg) self._ESPEAK_LIB = backend self._ESPEAK_VER = get_espeakng_version() if backend == "espeak-ng" else get_espeak_version() def auto_set_espeak_lib(self) -> None: - if is_tool("espeak-ng"): + if _is_tool("espeak-ng"): self._ESPEAK_LIB = "espeak-ng" self._ESPEAK_VER = get_espeakng_version() - elif is_tool("espeak"): + elif _is_tool("espeak"): self._ESPEAK_LIB = "espeak" self._ESPEAK_VER = get_espeak_version() else: - raise Exception("Cannot set backend automatically. espeak-ng or espeak not found") + msg = "Cannot set backend automatically. espeak-ng or espeak not found" + raise FileNotFoundError(msg) @staticmethod - def name(): + def name() -> str: return "espeak" - def phonemize_espeak(self, text: str, separator: str = "|", tie=False) -> str: + def phonemize_espeak(self, text: str, separator: str = "|", *, tie: bool = False) -> str: """Convert input text to phonemes. Args: @@ -185,12 +186,15 @@ def phonemize_espeak(self, text: str, separator: str = "|", tie=False) -> str: if tie: args.append("--tie=%s" % tie) - args.append(text) + tmp = tempfile.NamedTemporaryFile(mode="w+t", delete=False, encoding="utf8") + tmp.write(text) + tmp.close() + args.append("-f") + args.append(tmp.name) + # compute phonemes phonemes = "" - for line in _espeak_exe(self._ESPEAK_LIB, args, sync=True): - logging.debug("line: %s", repr(line)) - ph_decoded = line.decode("utf8").strip() + for line in _espeak_exe(self.backend, args): # espeak: # version 1.48.15: " p_Éš_ˈaÉĒ_ɚ t_ə n_oʊ_v_ˈɛ_m_b_ɚ t_w_ˈɛ_n_t_i t_ˈuː\n" # espeak-ng: @@ -200,16 +204,17 @@ def phonemize_espeak(self, text: str, separator: str = "|", tie=False) -> str: # "sɛʁtˈɛĖƒ mˈo kɔm (en)fˈʊtbɔːl(fr) ʒenˈɛʁ de- flˈaÉĄ də- lˈɑĖƒÉĄ." # phonemize needs to remove the language flags of the returned text: # "sɛʁtˈɛĖƒ mˈo kɔm fˈʊtbɔːl ʒenˈɛʁ de- flˈaÉĄ də- lˈɑĖƒÉĄ." - ph_decoded = re.sub(r"\(.+?\)", "", ph_decoded) + ph_decoded = re.sub(r"\(.+?\)", "", line) phonemes += ph_decoded.strip() + Path(tmp.name).unlink() return phonemes.replace("_", separator) - def _phonemize(self, text, separator=None): + def _phonemize(self, text: str, separator: str = "") -> str: return self.phonemize_espeak(text, separator, tie=False) @staticmethod - def supported_languages() -> Dict: + def supported_languages() -> dict[str, str]: """Get a dictionary of supported languages. Returns: @@ -219,16 +224,12 @@ def supported_languages() -> Dict: return {} args = ["--voices"] langs = {} - count = 0 - for line in _espeak_exe(_DEF_ESPEAK_LIB, args, sync=True): - line = line.decode("utf8").strip() + for count, line in enumerate(_espeak_exe(_DEF_ESPEAK_LIB, args)): if count > 0: cols = line.split() lang_code = cols[1] lang_name = cols[3] langs[lang_code] = lang_name - logging.debug("line: %s", repr(line)) - count += 1 return langs def version(self) -> str: @@ -237,16 +238,12 @@ def version(self) -> str: Returns: str: Version of the used backend. """ - args = ["--version"] - for line in _espeak_exe(self.backend, args, sync=True): - version = line.decode("utf8").strip().split()[2] - logging.debug("line: %s", repr(line)) - return version + return self.backend_version @classmethod - def is_available(cls): - """Return true if ESpeak is available else false""" - return is_tool("espeak") or is_tool("espeak-ng") + def is_available(cls) -> bool: + """Return true if ESpeak is available else false.""" + return _is_tool("espeak") or _is_tool("espeak-ng") if __name__ == "__main__": diff --git a/TTS/tts/utils/text/phonemizers/multi_phonemizer.py b/TTS/tts/utils/text/phonemizers/multi_phonemizer.py index 62a9c39322..1a9e98b091 100644 --- a/TTS/tts/utils/text/phonemizers/multi_phonemizer.py +++ b/TTS/tts/utils/text/phonemizers/multi_phonemizer.py @@ -1,7 +1,10 @@ +import logging from typing import Dict, List from TTS.tts.utils.text.phonemizers import DEF_LANG_TO_PHONEMIZER, get_phonemizer_by_name +logger = logging.getLogger(__name__) + class MultiPhonemizer: """🐸TTS multi-phonemizer that operates phonemizers for multiple langugages @@ -46,8 +49,8 @@ def supported_languages(self) -> List: def print_logs(self, level: int = 0): indent = "\t" * level - print(f"{indent}| > phoneme language: {self.supported_languages()}") - print(f"{indent}| > phoneme backend: {self.name()}") + logger.info("%s| phoneme language: %s", indent, self.supported_languages()) + logger.info("%s| phoneme backend: %s", indent, self.name()) # if __name__ == "__main__": diff --git a/TTS/tts/utils/text/tokenizer.py b/TTS/tts/utils/text/tokenizer.py index b7faf86e8a..f653cdf13f 100644 --- a/TTS/tts/utils/text/tokenizer.py +++ b/TTS/tts/utils/text/tokenizer.py @@ -1,3 +1,4 @@ +import logging from typing import Callable, Dict, List, Union from TTS.tts.utils.text import cleaners @@ -6,6 +7,8 @@ from TTS.tts.utils.text.phonemizers.multi_phonemizer import MultiPhonemizer from TTS.utils.generic_utils import get_import_path, import_class +logger = logging.getLogger(__name__) + class TTSTokenizer: """🐸TTS tokenizer to convert input characters to token IDs and back. @@ -73,8 +76,8 @@ def encode(self, text: str) -> List[int]: # discard but store not found characters if char not in self.not_found_characters: self.not_found_characters.append(char) - print(text) - print(f" [!] Character {repr(char)} not found in the vocabulary. Discarding it.") + logger.warning(text) + logger.warning("Character %s not found in the vocabulary. Discarding it.", repr(char)) return token_ids def decode(self, token_ids: List[int]) -> str: @@ -104,10 +107,13 @@ def text_to_ids(self, text: str, language: str = None) -> List[int]: # pylint: 5. Text to token IDs """ # TODO: text cleaner should pick the right routine based on the language + logger.debug("Tokenizer input text: %s", text) if self.text_cleaner is not None: text = self.text_cleaner(text) + logger.debug("Cleaned text: %s", text) if self.use_phonemes: text = self.phonemizer.phonemize(text, separator="", language=language) + logger.debug("Phonemes: %s", text) text = self.encode(text) if self.add_blank: text = self.intersperse_blank_char(text, True) @@ -135,16 +141,16 @@ def intersperse_blank_char(self, char_sequence: List[str], use_blank_char: bool def print_logs(self, level: int = 0): indent = "\t" * level - print(f"{indent}| > add_blank: {self.add_blank}") - print(f"{indent}| > use_eos_bos: {self.use_eos_bos}") - print(f"{indent}| > use_phonemes: {self.use_phonemes}") + logger.info("%s| add_blank: %s", indent, self.add_blank) + logger.info("%s| use_eos_bos: %s", indent, self.use_eos_bos) + logger.info("%s| use_phonemes: %s", indent, self.use_phonemes) if self.use_phonemes: - print(f"{indent}| > phonemizer:") + logger.info("%s| phonemizer:", indent) self.phonemizer.print_logs(level + 1) if len(self.not_found_characters) > 0: - print(f"{indent}| > {len(self.not_found_characters)} not found characters:") + logger.info("%s| %d characters not found:", indent, len(self.not_found_characters)) for char in self.not_found_characters: - print(f"{indent}| > {char}") + logger.info("%s| %s", indent, char) @staticmethod def init_from_config(config: "Coqpit", characters: "BaseCharacters" = None): diff --git a/TTS/utils/audio/numpy_transforms.py b/TTS/utils/audio/numpy_transforms.py index af88569fc3..0cba7fc8a8 100644 --- a/TTS/utils/audio/numpy_transforms.py +++ b/TTS/utils/audio/numpy_transforms.py @@ -1,5 +1,7 @@ +import logging +import os from io import BytesIO -from typing import Tuple +from typing import Any, Optional, Union import librosa import numpy as np @@ -7,17 +9,19 @@ import soundfile as sf from librosa import magphase, pyin +logger = logging.getLogger(__name__) + # For using kwargs # pylint: disable=unused-argument def build_mel_basis( *, - sample_rate: int = None, - fft_size: int = None, - num_mels: int = None, - mel_fmax: int = None, - mel_fmin: int = None, + sample_rate: int, + fft_size: int, + num_mels: int, + mel_fmin: int, + mel_fmax: Optional[int] = None, **kwargs, ) -> np.ndarray: """Build melspectrogram basis. @@ -31,9 +35,7 @@ def build_mel_basis( return librosa.filters.mel(sr=sample_rate, n_fft=fft_size, n_mels=num_mels, fmin=mel_fmin, fmax=mel_fmax) -def millisec_to_length( - *, frame_length_ms: int = None, frame_shift_ms: int = None, sample_rate: int = None, **kwargs -) -> Tuple[int, int]: +def millisec_to_length(*, frame_length_ms: float, frame_shift_ms: float, sample_rate: int, **kwargs) -> tuple[int, int]: """Compute hop and window length from milliseconds. Returns: @@ -58,7 +60,7 @@ def _exp(x, base): return np.exp(x) -def amp_to_db(*, x: np.ndarray = None, gain: float = 1, base: int = 10, **kwargs) -> np.ndarray: +def amp_to_db(*, x: np.ndarray, gain: float = 1, base: float = 10, **kwargs) -> np.ndarray: """Convert amplitude values to decibels. Args: @@ -74,7 +76,7 @@ def amp_to_db(*, x: np.ndarray = None, gain: float = 1, base: int = 10, **kwargs # pylint: disable=no-self-use -def db_to_amp(*, x: np.ndarray = None, gain: float = 1, base: int = 10, **kwargs) -> np.ndarray: +def db_to_amp(*, x: np.ndarray, gain: float = 1, base: float = 10, **kwargs) -> np.ndarray: """Convert decibels spectrogram to amplitude spectrogram. Args: @@ -101,18 +103,20 @@ def preemphasis(*, x: np.ndarray, coef: float = 0.97, **kwargs) -> np.ndarray: np.ndarray: Decorrelated audio signal. """ if coef == 0: - raise RuntimeError(" [!] Preemphasis is set 0.0.") + msg = " [!] Preemphasis is set 0.0." + raise RuntimeError(msg) return scipy.signal.lfilter([1, -coef], [1], x) -def deemphasis(*, x: np.ndarray = None, coef: float = 0.97, **kwargs) -> np.ndarray: +def deemphasis(*, x: np.ndarray, coef: float = 0.97, **kwargs) -> np.ndarray: """Reverse pre-emphasis.""" if coef == 0: - raise RuntimeError(" [!] Preemphasis is set 0.0.") + msg = " [!] Preemphasis is set 0.0." + raise ValueError(msg) return scipy.signal.lfilter([1], [1, -coef], x) -def spec_to_mel(*, spec: np.ndarray, mel_basis: np.ndarray = None, **kwargs) -> np.ndarray: +def spec_to_mel(*, spec: np.ndarray, mel_basis: np.ndarray, **kwargs) -> np.ndarray: """Convert a full scale linear spectrogram output of a network to a melspectrogram. Args: @@ -127,14 +131,14 @@ def spec_to_mel(*, spec: np.ndarray, mel_basis: np.ndarray = None, **kwargs) -> return np.dot(mel_basis, spec) -def mel_to_spec(*, mel: np.ndarray = None, mel_basis: np.ndarray = None, **kwargs) -> np.ndarray: +def mel_to_spec(*, mel: np.ndarray, mel_basis: np.ndarray, **kwargs) -> np.ndarray: """Convert a melspectrogram to full scale spectrogram.""" assert (mel < 0).sum() == 0, " [!] Input values must be non-negative." inv_mel_basis = np.linalg.pinv(mel_basis) return np.maximum(1e-10, np.dot(inv_mel_basis, mel)) -def wav_to_spec(*, wav: np.ndarray = None, **kwargs) -> np.ndarray: +def wav_to_spec(*, wav: np.ndarray, **kwargs) -> np.ndarray: """Compute a spectrogram from a waveform. Args: @@ -148,7 +152,7 @@ def wav_to_spec(*, wav: np.ndarray = None, **kwargs) -> np.ndarray: return S.astype(np.float32) -def wav_to_mel(*, wav: np.ndarray = None, mel_basis=None, **kwargs) -> np.ndarray: +def wav_to_mel(*, wav: np.ndarray, mel_basis: np.ndarray, **kwargs) -> np.ndarray: """Compute a melspectrogram from a waveform.""" D = stft(y=wav, **kwargs) S = spec_to_mel(spec=np.abs(D), mel_basis=mel_basis, **kwargs) @@ -161,20 +165,20 @@ def spec_to_wav(*, spec: np.ndarray, power: float = 1.5, **kwargs) -> np.ndarray return griffin_lim(spec=S**power, **kwargs) -def mel_to_wav(*, mel: np.ndarray = None, power: float = 1.5, **kwargs) -> np.ndarray: +def mel_to_wav(*, mel: np.ndarray, mel_basis: np.ndarray, power: float = 1.5, **kwargs) -> np.ndarray: """Convert a melspectrogram to a waveform using Griffi-Lim vocoder.""" S = mel.copy() - S = mel_to_spec(mel=S, mel_basis=kwargs["mel_basis"]) # Convert back to linear + S = mel_to_spec(mel=S, mel_basis=mel_basis) # Convert back to linear return griffin_lim(spec=S**power, **kwargs) ### STFT and ISTFT ### def stft( *, - y: np.ndarray = None, - fft_size: int = None, - hop_length: int = None, - win_length: int = None, + y: np.ndarray, + fft_size: int, + hop_length: Optional[int] = None, + win_length: Optional[int] = None, pad_mode: str = "reflect", window: str = "hann", center: bool = True, @@ -200,9 +204,9 @@ def stft( def istft( *, - y: np.ndarray = None, - hop_length: int = None, - win_length: int = None, + y: np.ndarray, + hop_length: Optional[int] = None, + win_length: Optional[int] = None, window: str = "hann", center: bool = True, **kwargs, @@ -217,12 +221,12 @@ def istft( return librosa.istft(y, hop_length=hop_length, win_length=win_length, center=center, window=window) -def griffin_lim(*, spec: np.ndarray = None, num_iter=60, **kwargs) -> np.ndarray: +def griffin_lim(*, spec: np.ndarray, num_iter=60, **kwargs) -> np.ndarray: angles = np.exp(2j * np.pi * np.random.rand(*spec.shape)) S_complex = np.abs(spec).astype(complex) y = istft(y=S_complex * angles, **kwargs) if not np.isfinite(y).all(): - print(" [!] Waveform is not finite everywhere. Skipping the GL.") + logger.warning("Waveform is not finite everywhere. Skipping the GL.") return np.array([0.0]) for _ in range(num_iter): angles = np.exp(1j * np.angle(stft(y=y, **kwargs))) @@ -230,11 +234,11 @@ def griffin_lim(*, spec: np.ndarray = None, num_iter=60, **kwargs) -> np.ndarray return y -def compute_stft_paddings( - *, x: np.ndarray = None, hop_length: int = None, pad_two_sides: bool = False, **kwargs -) -> Tuple[int, int]: - """Compute paddings used by Librosa's STFT. Compute right padding (final frame) or both sides padding - (first and final frames)""" +def compute_stft_paddings(*, x: np.ndarray, hop_length: int, pad_two_sides: bool = False, **kwargs) -> tuple[int, int]: + """Compute paddings used by Librosa's STFT. + + Compute right padding (final frame) or both sides padding (first and final frames). + """ pad = (x.shape[0] // hop_length + 1) * hop_length - x.shape[0] if not pad_two_sides: return 0, pad @@ -243,12 +247,12 @@ def compute_stft_paddings( def compute_f0( *, - x: np.ndarray = None, - pitch_fmax: float = None, - pitch_fmin: float = None, - hop_length: int = None, - win_length: int = None, - sample_rate: int = None, + x: np.ndarray, + pitch_fmax: Optional[float] = None, + pitch_fmin: Optional[float] = None, + hop_length: int, + win_length: int, + sample_rate: int, stft_pad_mode: str = "reflect", center: bool = True, **kwargs, @@ -320,19 +324,18 @@ def compute_energy(y: np.ndarray, **kwargs) -> np.ndarray: """ x = stft(y=y, **kwargs) mag, _ = magphase(x) - energy = np.sqrt(np.sum(mag**2, axis=0)) - return energy + return np.sqrt(np.sum(mag**2, axis=0)) ### Audio Processing ### def find_endpoint( *, - wav: np.ndarray = None, + wav: np.ndarray, trim_db: float = -40, - sample_rate: int = None, - min_silence_sec=0.8, - gain: float = None, - base: int = None, + sample_rate: int, + min_silence_sec: float = 0.8, + gain: float = 1, + base: float = 10, **kwargs, ) -> int: """Find the last point without silence at the end of a audio signal. @@ -341,8 +344,8 @@ def find_endpoint( wav (np.ndarray): Audio signal. threshold_db (int, optional): Silence threshold in decibels. Defaults to -40. min_silence_sec (float, optional): Ignore silences that are shorter then this in secs. Defaults to 0.8. - gian (float, optional): Gain to be used to convert trim_db to trim_amp. Defaults to None. - base (int, optional): Base of the logarithm used to convert trim_db to trim_amp. Defaults to 10. + gain (float, optional): Gain factor to be used to convert trim_db to trim_amp. Defaults to 1. + base (float, optional): Base of the logarithm used to convert trim_db to trim_amp. Defaults to 10. Returns: int: Last point without silence. @@ -358,20 +361,20 @@ def find_endpoint( def trim_silence( *, - wav: np.ndarray = None, - sample_rate: int = None, - trim_db: float = None, - win_length: int = None, - hop_length: int = None, + wav: np.ndarray, + sample_rate: int, + trim_db: float = 60, + win_length: int, + hop_length: int, **kwargs, ) -> np.ndarray: - """Trim silent parts with a threshold and 0.01 sec margin""" + """Trim silent parts with a threshold and 0.01 sec margin.""" margin = int(sample_rate * 0.01) wav = wav[margin:-margin] return librosa.effects.trim(wav, top_db=trim_db, frame_length=win_length, hop_length=hop_length)[0] -def volume_norm(*, x: np.ndarray = None, coef: float = 0.95, **kwargs) -> np.ndarray: +def volume_norm(*, x: np.ndarray, coef: float = 0.95, **kwargs) -> np.ndarray: """Normalize the volume of an audio signal. Args: @@ -384,7 +387,7 @@ def volume_norm(*, x: np.ndarray = None, coef: float = 0.95, **kwargs) -> np.nda return x / abs(x).max() * coef -def rms_norm(*, wav: np.ndarray = None, db_level: float = -27.0, **kwargs) -> np.ndarray: +def rms_norm(*, wav: np.ndarray, db_level: float = -27.0, **kwargs) -> np.ndarray: r = 10 ** (db_level / 20) a = np.sqrt((len(wav) * (r**2)) / np.sum(wav**2)) return wav * a @@ -401,11 +404,12 @@ def rms_volume_norm(*, x: np.ndarray, db_level: float = -27.0, **kwargs) -> np.n np.ndarray: RMS normalized waveform. """ assert -99 <= db_level <= 0, " [!] db_level should be between -99 and 0" - wav = rms_norm(wav=x, db_level=db_level) - return wav + return rms_norm(wav=x, db_level=db_level) -def load_wav(*, filename: str, sample_rate: int = None, resample: bool = False, **kwargs) -> np.ndarray: +def load_wav( + *, filename: Union[str, os.PathLike[Any]], sample_rate: Optional[int] = None, resample: bool = False, **kwargs +) -> np.ndarray: """Read a wav file using Librosa and optionally resample, silence trim, volume normalize. Resampling slows down loading the file significantly. Therefore it is recommended to resample the file before. @@ -424,19 +428,39 @@ def load_wav(*, filename: str, sample_rate: int = None, resample: bool = False, else: # SF is faster than librosa for loading files x, _ = sf.read(filename) + if x.ndim != 1: + logger.warning("Found multi-channel audio. Converting to mono: %s", filename) + x = librosa.to_mono(x) return x -def save_wav(*, wav: np.ndarray, path: str, sample_rate: int = None, pipe_out=None, **kwargs) -> None: +def save_wav( + *, + wav: np.ndarray, + path: Union[str, os.PathLike[Any]], + sample_rate: int, + pipe_out=None, + do_rms_norm: bool = False, + db_level: float = -27.0, + **kwargs, +) -> None: """Save float waveform to a file using Scipy. Args: wav (np.ndarray): Waveform with float values in range [-1, 1] to save. path (str): Path to a output file. - sr (int, optional): Sampling rate used for saving to the file. Defaults to None. + sr (int): Sampling rate used for saving to the file. Defaults to None. pipe_out (BytesIO, optional): Flag to stdout the generated TTS wav file for shell pipe. + do_rms_norm (bool): Whether to apply RMS normalization + db_level (float): Target dB level in RMS. """ - wav_norm = wav * (32767 / max(0.01, np.max(np.abs(wav)))) + if do_rms_norm: + if db_level is None: + msg = "`db_level` cannot be None with `do_rms_norm=True`" + raise ValueError(msg) + wav_norm = rms_volume_norm(x=wav, db_level=db_level) + else: + wav_norm = wav * (32767 / max(0.01, np.max(np.abs(wav)))) wav_norm = wav_norm.astype(np.int16) if pipe_out: @@ -459,8 +483,7 @@ def mulaw_encode(*, wav: np.ndarray, mulaw_qc: int, **kwargs) -> np.ndarray: def mulaw_decode(*, wav, mulaw_qc: int, **kwargs) -> np.ndarray: """Recovers waveform from quantized values.""" mu = 2**mulaw_qc - 1 - x = np.sign(wav) / mu * ((1 + mu) ** np.abs(wav) - 1) - return x + return np.sign(wav) / mu * ((1 + mu) ** np.abs(wav) - 1) def encode_16bits(*, x: np.ndarray, **kwargs) -> np.ndarray: diff --git a/TTS/utils/audio/processor.py b/TTS/utils/audio/processor.py index c53bad562e..bf07333aea 100644 --- a/TTS/utils/audio/processor.py +++ b/TTS/utils/audio/processor.py @@ -1,10 +1,9 @@ -from io import BytesIO -from typing import Dict, Tuple +import logging +import os +from typing import Any, Optional, Union import librosa import numpy as np -import scipy.io.wavfile -import scipy.signal from TTS.tts.utils.helpers import StandardScaler from TTS.utils.audio.numpy_transforms import ( @@ -20,16 +19,19 @@ millisec_to_length, preemphasis, rms_volume_norm, + save_wav, spec_to_mel, stft, trim_silence, volume_norm, ) +logger = logging.getLogger(__name__) + # pylint: disable=too-many-public-methods -class AudioProcessor(object): +class AudioProcessor: """Audio Processor for TTS. Note: @@ -132,10 +134,6 @@ class AudioProcessor(object): stats_path (str, optional): Path to the computed stats file. Defaults to None. - - verbose (bool, optional): - enable/disable logging. Defaults to True. - """ def __init__( @@ -172,9 +170,8 @@ def __init__( do_rms_norm=False, db_level=None, stats_path=None, - verbose=True, **_, - ): + ) -> None: # setup class attributed self.sample_rate = sample_rate self.resample = resample @@ -212,7 +209,8 @@ def __init__( elif log_func == "np.log10": self.base = 10 else: - raise ValueError(" [!] unknown `log_func` value.") + msg = " [!] unknown `log_func` value." + raise ValueError(msg) # setup stft parameters if hop_length is None: # compute stft parameters from given time values @@ -228,10 +226,9 @@ def __init__( self.win_length <= self.fft_size ), f" [!] win_length cannot be larger than fft_size - {self.win_length} vs {self.fft_size}" members = vars(self) - if verbose: - print(" > Setting up Audio Processor...") - for key, value in members.items(): - print(" | > {}:{}".format(key, value)) + logger.info("Setting up Audio Processor...") + for key, value in members.items(): + logger.info(" | %s: %s", key, value) # create spectrogram utils self.mel_basis = build_mel_basis( sample_rate=self.sample_rate, @@ -250,14 +247,14 @@ def __init__( self.symmetric_norm = None @staticmethod - def init_from_config(config: "Coqpit", verbose=True): + def init_from_config(config: "Coqpit"): if "audio" in config: - return AudioProcessor(verbose=verbose, **config.audio) - return AudioProcessor(verbose=verbose, **config) + return AudioProcessor(**config.audio) + return AudioProcessor(**config) ### normalization ### def normalize(self, S: np.ndarray) -> np.ndarray: - """Normalize values into `[0, self.max_norm]` or `[-self.max_norm, self.max_norm]` + """Normalize values into `[0, self.max_norm]` or `[-self.max_norm, self.max_norm]`. Args: S (np.ndarray): Spectrogram to normalize. @@ -275,10 +272,10 @@ def normalize(self, S: np.ndarray) -> np.ndarray: if hasattr(self, "mel_scaler"): if S.shape[0] == self.num_mels: return self.mel_scaler.transform(S.T).T - elif S.shape[0] == self.fft_size / 2: + if S.shape[0] == self.fft_size / 2: return self.linear_scaler.transform(S.T).T - else: - raise RuntimeError(" [!] Mean-Var stats does not match the given feature dimensions.") + msg = " [!] Mean-Var stats does not match the given feature dimensions." + raise RuntimeError(msg) # range normalization S -= self.ref_level_db # discard certain range of DB assuming it is air noise S_norm = (S - self.min_level_db) / (-self.min_level_db) @@ -289,13 +286,11 @@ def normalize(self, S: np.ndarray) -> np.ndarray: S_norm, -self.max_norm, self.max_norm # pylint: disable=invalid-unary-operand-type ) return S_norm - else: - S_norm = self.max_norm * S_norm - if self.clip_norm: - S_norm = np.clip(S_norm, 0, self.max_norm) - return S_norm - else: - return S + S_norm = self.max_norm * S_norm + if self.clip_norm: + S_norm = np.clip(S_norm, 0, self.max_norm) + return S_norm + return S def denormalize(self, S: np.ndarray) -> np.ndarray: """Denormalize spectrogram values. @@ -316,10 +311,10 @@ def denormalize(self, S: np.ndarray) -> np.ndarray: if hasattr(self, "mel_scaler"): if S_denorm.shape[0] == self.num_mels: return self.mel_scaler.inverse_transform(S_denorm.T).T - elif S_denorm.shape[0] == self.fft_size / 2: + if S_denorm.shape[0] == self.fft_size / 2: return self.linear_scaler.inverse_transform(S_denorm.T).T - else: - raise RuntimeError(" [!] Mean-Var stats does not match the given feature dimensions.") + msg = " [!] Mean-Var stats does not match the given feature dimensions." + raise RuntimeError(msg) if self.symmetric_norm: if self.clip_norm: S_denorm = np.clip( @@ -327,16 +322,14 @@ def denormalize(self, S: np.ndarray) -> np.ndarray: ) S_denorm = ((S_denorm + self.max_norm) * -self.min_level_db / (2 * self.max_norm)) + self.min_level_db return S_denorm + self.ref_level_db - else: - if self.clip_norm: - S_denorm = np.clip(S_denorm, 0, self.max_norm) - S_denorm = (S_denorm * -self.min_level_db / self.max_norm) + self.min_level_db - return S_denorm + self.ref_level_db - else: - return S_denorm + if self.clip_norm: + S_denorm = np.clip(S_denorm, 0, self.max_norm) + S_denorm = (S_denorm * -self.min_level_db / self.max_norm) + self.min_level_db + return S_denorm + self.ref_level_db + return S_denorm ### Mean-STD scaling ### - def load_stats(self, stats_path: str) -> Tuple[np.array, np.array, np.array, np.array, Dict]: + def load_stats(self, stats_path: str) -> tuple[np.array, np.array, np.array, np.array, dict]: """Loading mean and variance statistics from a `npy` file. Args: @@ -354,7 +347,7 @@ def load_stats(self, stats_path: str) -> Tuple[np.array, np.array, np.array, np. stats_config = stats["audio_config"] # check all audio parameters used for computing stats skip_parameters = ["griffin_lim_iters", "stats_path", "do_trim_silence", "ref_level_db", "power"] - for key in stats_config.keys(): + for key in stats_config: if key in skip_parameters: continue if key not in ["sample_rate", "trim_db"]: @@ -418,10 +411,7 @@ def spectrogram(self, y: np.ndarray) -> np.ndarray: win_length=self.win_length, pad_mode=self.stft_pad_mode, ) - if self.do_amp_to_db_linear: - S = amp_to_db(x=np.abs(D), gain=self.spec_gain, base=self.base) - else: - S = np.abs(D) + S = amp_to_db(x=np.abs(D), gain=self.spec_gain, base=self.base) if self.do_amp_to_db_linear else np.abs(D) return self.normalize(S).astype(np.float32) def melspectrogram(self, y: np.ndarray) -> np.ndarray: @@ -470,8 +460,7 @@ def out_linear_to_mel(self, linear_spec: np.ndarray) -> np.ndarray: S = db_to_amp(x=S, gain=self.spec_gain, base=self.base) S = spec_to_mel(spec=np.abs(S), mel_basis=self.mel_basis) S = amp_to_db(x=S, gain=self.spec_gain, base=self.base) - mel = self.normalize(S) - return mel + return self.normalize(S) def _griffin_lim(self, S): return griffin_lim( @@ -505,7 +494,7 @@ def compute_f0(self, x: np.ndarray) -> np.ndarray: if len(x) % self.hop_length == 0: x = np.pad(x, (0, self.hop_length // 2), mode=self.stft_pad_mode) - f0 = compute_f0( + return compute_f0( x=x, pitch_fmax=self.pitch_fmax, pitch_fmin=self.pitch_fmin, @@ -516,8 +505,6 @@ def compute_f0(self, x: np.ndarray) -> np.ndarray: center=True, ) - return f0 - ### Audio Processing ### def find_endpoint(self, wav: np.ndarray, min_silence_sec=0.8) -> int: """Find the last point without silence at the end of a audio signal. @@ -540,7 +527,7 @@ def find_endpoint(self, wav: np.ndarray, min_silence_sec=0.8) -> int: ) def trim_silence(self, wav): - """Trim silent parts with a threshold and 0.01 sec margin""" + """Trim silent parts with a threshold and 0.01 sec margin.""" return trim_silence( wav=wav, sample_rate=self.sample_rate, @@ -561,21 +548,8 @@ def sound_norm(x: np.ndarray) -> np.ndarray: """ return volume_norm(x=x) - def rms_volume_norm(self, x: np.ndarray, db_level: float = None) -> np.ndarray: - """Normalize the volume based on RMS of the signal. - - Args: - x (np.ndarray): Raw waveform. - - Returns: - np.ndarray: RMS normalized waveform. - """ - if db_level is None: - db_level = self.db_level - return rms_volume_norm(x=x, db_level=db_level) - ### save and load ### - def load_wav(self, filename: str, sr: int = None) -> np.ndarray: + def load_wav(self, filename: Union[str, os.PathLike[Any]], sr: Optional[int] = None) -> np.ndarray: """Read a wav file using Librosa and optionally resample, silence trim, volume normalize. Resampling slows down loading the file significantly. Therefore it is recommended to resample the file before. @@ -595,14 +569,16 @@ def load_wav(self, filename: str, sr: int = None) -> np.ndarray: try: x = self.trim_silence(x) except ValueError: - print(f" [!] File cannot be trimmed for silence - {filename}") + logger.exception("File cannot be trimmed for silence - %s", filename) if self.do_sound_norm: x = self.sound_norm(x) if self.do_rms_norm: - x = self.rms_volume_norm(x, self.db_level) + x = rms_volume_norm(x=x, db_level=self.db_level) return x - def save_wav(self, wav: np.ndarray, path: str, sr: int = None, pipe_out=None) -> None: + def save_wav( + self, wav: np.ndarray, path: Union[str, os.PathLike[Any]], sr: Optional[int] = None, pipe_out=None + ) -> None: """Save a waveform to a file using Scipy. Args: @@ -611,18 +587,14 @@ def save_wav(self, wav: np.ndarray, path: str, sr: int = None, pipe_out=None) -> sr (int, optional): Sampling rate used for saving to the file. Defaults to None. pipe_out (BytesIO, optional): Flag to stdout the generated TTS wav file for shell pipe. """ - if self.do_rms_norm: - wav_norm = self.rms_volume_norm(wav, self.db_level) * 32767 - else: - wav_norm = wav * (32767 / max(0.01, np.max(np.abs(wav)))) - - wav_norm = wav_norm.astype(np.int16) - if pipe_out: - wav_buffer = BytesIO() - scipy.io.wavfile.write(wav_buffer, sr if sr else self.sample_rate, wav_norm) - wav_buffer.seek(0) - pipe_out.buffer.write(wav_buffer.read()) - scipy.io.wavfile.write(path, sr if sr else self.sample_rate, wav_norm) + save_wav( + wav=wav, + path=path, + sample_rate=sr if sr else self.sample_rate, + pipe_out=pipe_out, + do_rms_norm=self.do_rms_norm, + db_level=self.db_level, + ) def get_duration(self, filename: str) -> float: """Get the duration of a wav file using Librosa. diff --git a/TTS/utils/audio/torch_transforms.py b/TTS/utils/audio/torch_transforms.py index fd40ebb048..59bb23cc4f 100644 --- a/TTS/utils/audio/torch_transforms.py +++ b/TTS/utils/audio/torch_transforms.py @@ -1,7 +1,113 @@ +import logging + import librosa import torch from torch import nn +logger = logging.getLogger(__name__) + + +hann_window = {} +mel_basis = {} + + +def amp_to_db(x: torch.Tensor, *, spec_gain: float = 1.0, clip_val: float = 1e-5) -> torch.Tensor: + """Spectral normalization / dynamic range compression.""" + return torch.log(torch.clamp(x, min=clip_val) * spec_gain) + + +def db_to_amp(x: torch.Tensor, *, spec_gain: float = 1.0) -> torch.Tensor: + """Spectral denormalization / dynamic range decompression.""" + return torch.exp(x) / spec_gain + + +def wav_to_spec(y: torch.Tensor, n_fft: int, hop_length: int, win_length: int, *, center: bool = False) -> torch.Tensor: + """ + Args Shapes: + - y : :math:`[B, 1, T]` + + Return Shapes: + - spec : :math:`[B,C,T]` + """ + y = y.squeeze(1) + + if torch.min(y) < -1.0: + logger.info("min value is %.3f", torch.min(y)) + if torch.max(y) > 1.0: + logger.info("max value is %.3f", torch.max(y)) + + global hann_window + wnsize_dtype_device = f"{win_length}_{y.dtype}_{y.device}" + if wnsize_dtype_device not in hann_window: + hann_window[wnsize_dtype_device] = torch.hann_window(win_length).to(dtype=y.dtype, device=y.device) + + y = torch.nn.functional.pad( + y.unsqueeze(1), + (int((n_fft - hop_length) / 2), int((n_fft - hop_length) / 2)), + mode="reflect", + ) + y = y.squeeze(1) + + spec = torch.view_as_real( + torch.stft( + y, + n_fft, + hop_length=hop_length, + win_length=win_length, + window=hann_window[wnsize_dtype_device], + center=center, + pad_mode="reflect", + normalized=False, + onesided=True, + return_complex=True, + ) + ) + + return torch.sqrt(spec.pow(2).sum(-1) + 1e-6) + + +def spec_to_mel( + spec: torch.Tensor, n_fft: int, num_mels: int, sample_rate: int, fmin: float, fmax: float +) -> torch.Tensor: + """ + Args Shapes: + - spec : :math:`[B,C,T]` + + Return Shapes: + - mel : :math:`[B,C,T]` + """ + global mel_basis + fmax_dtype_device = f"{n_fft}_{fmax}_{spec.dtype}_{spec.device}" + if fmax_dtype_device not in mel_basis: + # TODO: switch librosa to torchaudio + mel = librosa.filters.mel(sr=sample_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax) + mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device) + mel = torch.matmul(mel_basis[fmax_dtype_device], spec) + return amp_to_db(mel) + + +def wav_to_mel( + y: torch.Tensor, + n_fft: int, + num_mels: int, + sample_rate: int, + hop_length: int, + win_length: int, + fmin: float, + fmax: float, + *, + center: bool = False, +) -> torch.Tensor: + """ + Args Shapes: + - y : :math:`[B, 1, T]` + + Return Shapes: + - spec : :math:`[B,C,T]` + """ + spec = wav_to_spec(y, n_fft, hop_length, win_length, center=center) + return spec_to_mel(spec, n_fft, num_mels, sample_rate, fmin, fmax) + class TorchSTFT(nn.Module): # pylint: disable=abstract-method """Some of the audio processing funtions using Torch for faster batch processing. @@ -119,17 +225,19 @@ def __call__(self, x): padding = int((self.n_fft - self.hop_length) / 2) x = torch.nn.functional.pad(x, (padding, padding), mode="reflect") # B x D x T x 2 - o = torch.stft( - x.squeeze(1), - self.n_fft, - self.hop_length, - self.win_length, - self.window, - center=True, - pad_mode="reflect", # compatible with audio.py - normalized=self.normalized, - onesided=True, - return_complex=False, + o = torch.view_as_real( + torch.stft( + x.squeeze(1), + self.n_fft, + self.hop_length, + self.win_length, + self.window, + center=True, + pad_mode="reflect", # compatible with audio.py + normalized=self.normalized, + onesided=True, + return_complex=True, + ) ) M = o[:, :, :, 0] P = o[:, :, :, 1] @@ -155,11 +263,3 @@ def _build_mel_basis(self): norm=self.mel_norm, ) self.mel_basis = torch.from_numpy(mel_basis).float() - - @staticmethod - def _amp_to_db(x, spec_gain=1.0): - return torch.log(torch.clamp(x, min=1e-5) * spec_gain) - - @staticmethod - def _db_to_amp(x, spec_gain=1.0): - return torch.exp(x) / spec_gain diff --git a/TTS/utils/download.py b/TTS/utils/download.py index 3f06b57824..e94b1d68c8 100644 --- a/TTS/utils/download.py +++ b/TTS/utils/download.py @@ -12,6 +12,8 @@ from torch.utils.model_zoo import tqdm +logger = logging.getLogger(__name__) + def stream_url( url: str, start_byte: Optional[int] = None, block_size: int = 32 * 1024, progress_bar: bool = True @@ -36,13 +38,16 @@ def stream_url( if start_byte: req.headers["Range"] = "bytes={}-".format(start_byte) - with urllib.request.urlopen(req) as upointer, tqdm( - unit="B", - unit_scale=True, - unit_divisor=1024, - total=url_size, - disable=not progress_bar, - ) as pbar: + with ( + urllib.request.urlopen(req) as upointer, + tqdm( + unit="B", + unit_scale=True, + unit_divisor=1024, + total=url_size, + disable=not progress_bar, + ) as pbar, + ): num_bytes = 0 while True: chunk = upointer.read(block_size) @@ -146,20 +151,20 @@ def extract_archive(from_path: str, to_path: Optional[str] = None, overwrite: bo Returns: list: List of paths to extracted files even if not overwritten. """ - + logger.info("Extracting archive file...") if to_path is None: to_path = os.path.dirname(from_path) try: with tarfile.open(from_path, "r") as tar: - logging.info("Opened tar file %s.", from_path) + logger.info("Opened tar file %s.", from_path) files = [] for file_ in tar: # type: Any file_path = os.path.join(to_path, file_.name) if file_.isfile(): files.append(file_path) if os.path.exists(file_path): - logging.info("%s already extracted.", file_path) + logger.info("%s already extracted.", file_path) if not overwrite: continue tar.extract(file_, to_path) @@ -169,12 +174,12 @@ def extract_archive(from_path: str, to_path: Optional[str] = None, overwrite: bo try: with zipfile.ZipFile(from_path, "r") as zfile: - logging.info("Opened zip file %s.", from_path) + logger.info("Opened zip file %s.", from_path) files = zfile.namelist() for file_ in files: file_path = os.path.join(to_path, file_) if os.path.exists(file_path): - logging.info("%s already extracted.", file_path) + logger.info("%s already extracted.", file_path) if not overwrite: continue zfile.extract(file_, to_path) @@ -198,9 +203,10 @@ def download_kaggle_dataset(dataset_path: str, dataset_name: str, output_path: s import kaggle # pylint: disable=import-outside-toplevel kaggle.api.authenticate() - print(f"""\nDownloading {dataset_name}...""") + logger.info("Downloading %s...", dataset_name) kaggle.api.dataset_download_files(dataset_path, path=data_path, unzip=True) except OSError: - print( - f"""[!] in order to download kaggle datasets, you need to have a kaggle api token stored in your {os.path.join(expanduser('~'), '.kaggle/kaggle.json')}""" + logger.exception( + "In order to download kaggle datasets, you need to have a kaggle api token stored in your %s", + os.path.join(expanduser("~"), ".kaggle/kaggle.json"), ) diff --git a/TTS/utils/downloaders.py b/TTS/utils/downloaders.py index 104dc7b94e..8705873982 100644 --- a/TTS/utils/downloaders.py +++ b/TTS/utils/downloaders.py @@ -1,8 +1,11 @@ +import logging import os from typing import Optional from TTS.utils.download import download_kaggle_dataset, download_url, extract_archive +logger = logging.getLogger(__name__) + def download_ljspeech(path: str): """Download and extract LJSpeech dataset @@ -15,7 +18,6 @@ def download_ljspeech(path: str): download_url(url, path) basename = os.path.basename(url) archive = os.path.join(path, basename) - print(" > Extracting archive file...") extract_archive(archive) @@ -35,7 +37,6 @@ def download_vctk(path: str, use_kaggle: Optional[bool] = False): download_url(url, path) basename = os.path.basename(url) archive = os.path.join(path, basename) - print(" > Extracting archive file...") extract_archive(archive) @@ -71,19 +72,17 @@ def download_libri_tts(path: str, subset: Optional[str] = "all"): os.makedirs(path, exist_ok=True) if subset == "all": for sub, val in subset_dict.items(): - print(f" > Downloading {sub}...") + logger.info("Downloading %s...", sub) download_url(val, path) basename = os.path.basename(val) archive = os.path.join(path, basename) - print(" > Extracting archive file...") extract_archive(archive) - print(" > All subsets downloaded") + logger.info("All subsets downloaded") else: url = subset_dict[subset] download_url(url, path) basename = os.path.basename(url) archive = os.path.join(path, basename) - print(" > Extracting archive file...") extract_archive(archive) @@ -98,7 +97,6 @@ def download_thorsten_de(path: str): download_url(url, path) basename = os.path.basename(url) archive = os.path.join(path, basename) - print(" > Extracting archive file...") extract_archive(archive) @@ -122,5 +120,4 @@ def download_mailabs(path: str, language: str = "english"): download_url(url, path) basename = os.path.basename(url) archive = os.path.join(path, basename) - print(" > Extracting archive file...") extract_archive(archive) diff --git a/TTS/utils/generic_utils.py b/TTS/utils/generic_utils.py index 4fa4741ab7..54bb5ba825 100644 --- a/TTS/utils/generic_utils.py +++ b/TTS/utils/generic_utils.py @@ -4,82 +4,26 @@ import logging import os import re -import subprocess -import sys from pathlib import Path -from typing import Dict +from typing import Any, Callable, Dict, Optional, TextIO, TypeVar, Union -import fsspec import torch +from packaging.version import Version +from typing_extensions import TypeIs +logger = logging.getLogger(__name__) -def to_cuda(x: torch.Tensor) -> torch.Tensor: - if x is None: - return None - if torch.is_tensor(x): - x = x.contiguous() - if torch.cuda.is_available(): - x = x.cuda(non_blocking=True) - return x - - -def get_cuda(): - use_cuda = torch.cuda.is_available() - device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") - return use_cuda, device - - -def get_git_branch(): - try: - out = subprocess.check_output(["git", "branch"]).decode("utf8") - current = next(line for line in out.split("\n") if line.startswith("*")) - current.replace("* ", "") - except subprocess.CalledProcessError: - current = "inside_docker" - except (FileNotFoundError, StopIteration) as e: - current = "unknown" - return current - - -def get_commit_hash(): - """https://stackoverflow.com/questions/14989858/get-the-current-git-hash-in-a-python-script""" - # try: - # subprocess.check_output(['git', 'diff-index', '--quiet', - # 'HEAD']) # Verify client is clean - # except: - # raise RuntimeError( - # " !! Commit before training to get the commit hash.") - try: - commit = subprocess.check_output(["git", "rev-parse", "--short", "HEAD"]).decode().strip() - # Not copying .git folder into docker container - except (subprocess.CalledProcessError, FileNotFoundError): - commit = "0000000" - return commit - - -def get_experiment_folder_path(root_path, model_name): - """Get an experiment folder path with the current date and time""" - date_str = datetime.datetime.now().strftime("%B-%d-%Y_%I+%M%p") - commit_hash = get_commit_hash() - output_folder = os.path.join(root_path, model_name + "-" + date_str + "-" + commit_hash) - return output_folder - - -def remove_experiment_folder(experiment_path): - """Check folder if there is a checkpoint, otherwise remove the folder""" - fs = fsspec.get_mapper(experiment_path).fs - checkpoint_files = fs.glob(experiment_path + "/*.pth") - if not checkpoint_files: - if fs.exists(experiment_path): - fs.rm(experiment_path, recursive=True) - print(" ! Run is removed from {}".format(experiment_path)) - else: - print(" ! Run is kept in {}".format(experiment_path)) - - -def count_parameters(model): - r"""Count number of trainable parameters in a network""" - return sum(p.numel() for p in model.parameters() if p.requires_grad) +_T = TypeVar("_T") + + +def exists(val: Union[_T, None]) -> TypeIs[_T]: + return val is not None + + +def default(val: Union[_T, None], d: Union[_T, Callable[[], _T]]) -> _T: + if exists(val): + return val + return d() if callable(d) else d def to_camel(text): @@ -124,47 +68,6 @@ def get_import_path(obj: object) -> str: return ".".join([type(obj).__module__, type(obj).__name__]) -def get_user_data_dir(appname): - TTS_HOME = os.environ.get("TTS_HOME") - XDG_DATA_HOME = os.environ.get("XDG_DATA_HOME") - if TTS_HOME is not None: - ans = Path(TTS_HOME).expanduser().resolve(strict=False) - elif XDG_DATA_HOME is not None: - ans = Path(XDG_DATA_HOME).expanduser().resolve(strict=False) - elif sys.platform == "win32": - import winreg # pylint: disable=import-outside-toplevel - - key = winreg.OpenKey( - winreg.HKEY_CURRENT_USER, r"Software\Microsoft\Windows\CurrentVersion\Explorer\Shell Folders" - ) - dir_, _ = winreg.QueryValueEx(key, "Local AppData") - ans = Path(dir_).resolve(strict=False) - elif sys.platform == "darwin": - ans = Path("~/Library/Application Support/").expanduser() - else: - ans = Path.home().joinpath(".local/share") - return ans.joinpath(appname) - - -def set_init_dict(model_dict, checkpoint_state, c): - # Partial initialization: if there is a mismatch with new and old layer, it is skipped. - for k, v in checkpoint_state.items(): - if k not in model_dict: - print(" | > Layer missing in the model definition: {}".format(k)) - # 1. filter out unnecessary keys - pretrained_dict = {k: v for k, v in checkpoint_state.items() if k in model_dict} - # 2. filter out different size layers - pretrained_dict = {k: v for k, v in pretrained_dict.items() if v.numel() == model_dict[k].numel()} - # 3. skip reinit layers - if c.has("reinit_layers") and c.reinit_layers is not None: - for reinit_layer_name in c.reinit_layers: - pretrained_dict = {k: v for k, v in pretrained_dict.items() if reinit_layer_name not in k} - # 4. overwrite entries in the existing state dict - model_dict.update(pretrained_dict) - print(" | > {} / {} layers are restored.".format(len(pretrained_dict), len(model_dict))) - return model_dict - - def format_aux_input(def_args: Dict, kwargs: Dict) -> Dict: """Format kwargs to hande auxilary inputs to models. @@ -182,58 +85,66 @@ def format_aux_input(def_args: Dict, kwargs: Dict) -> Dict: return kwargs -class KeepAverage: - def __init__(self): - self.avg_values = {} - self.iters = {} +def get_timestamp() -> str: + return datetime.datetime.now().strftime("%y%m%d-%H%M%S") - def __getitem__(self, key): - return self.avg_values[key] - def items(self): - return self.avg_values.items() +class ConsoleFormatter(logging.Formatter): + """Custom formatter that prints logging.INFO messages without the level name. - def add_value(self, name, init_val=0, init_iter=0): - self.avg_values[name] = init_val - self.iters[name] = init_iter + Source: https://stackoverflow.com/a/62488520 + """ - def update_value(self, name, value, weighted_avg=False): - if name not in self.avg_values: - # add value if not exist before - self.add_value(name, init_val=value) + def format(self, record): + if record.levelno == logging.INFO: + self._style._fmt = "%(message)s" else: - # else update existing value - if weighted_avg: - self.avg_values[name] = 0.99 * self.avg_values[name] + 0.01 * value - self.iters[name] += 1 - else: - self.avg_values[name] = self.avg_values[name] * self.iters[name] + value - self.iters[name] += 1 - self.avg_values[name] /= self.iters[name] - - def add_values(self, name_dict): - for key, value in name_dict.items(): - self.add_value(key, init_val=value) + self._style._fmt = "%(levelname)s: %(message)s" + return super().format(record) - def update_values(self, value_dict): - for key, value in value_dict.items(): - self.update_value(key, value) +def setup_logger( + logger_name: str, + level: int = logging.INFO, + *, + formatter: Optional[logging.Formatter] = None, + stream: Optional[TextIO] = None, + log_dir: Optional[Union[str, os.PathLike[Any]]] = None, + log_name: str = "log", +) -> None: + """Set up a logger. -def get_timestamp(): - return datetime.now().strftime("%y%m%d-%H%M%S") - - -def setup_logger(logger_name, root, phase, level=logging.INFO, screen=False, tofile=False): + Args: + logger_name: Name of the logger to set up + level: Logging level + formatter: Formatter for the logger + stream: Add a StreamHandler for the given stream, e.g. sys.stderr or sys.stdout + log_dir: Folder to write the log file (no file created if None) + log_name: Prefix of the log file name + """ lg = logging.getLogger(logger_name) - formatter = logging.Formatter("%(asctime)s.%(msecs)03d - %(levelname)s: %(message)s", datefmt="%y-%m-%d %H:%M:%S") + if formatter is None: + formatter = logging.Formatter( + "%(asctime)s.%(msecs)03d - %(levelname)-8s - %(name)s: %(message)s", datefmt="%y-%m-%d %H:%M:%S" + ) lg.setLevel(level) - if tofile: - log_file = os.path.join(root, phase + "_{}.log".format(get_timestamp())) + if log_dir is not None: + Path(log_dir).mkdir(exist_ok=True, parents=True) + log_file = Path(log_dir) / f"{log_name}_{get_timestamp()}.log" fh = logging.FileHandler(log_file, mode="w") fh.setFormatter(formatter) lg.addHandler(fh) - if screen: - sh = logging.StreamHandler() + if stream is not None: + sh = logging.StreamHandler(stream) sh.setFormatter(formatter) lg.addHandler(sh) + + +def is_pytorch_at_least_2_4() -> bool: + """Check if the installed Pytorch version is 2.4 or higher.""" + return Version(torch.__version__) >= Version("2.4") + + +def optional_to_str(x: Optional[Any]) -> str: + """Convert input to string, using empty string if input is None.""" + return "" if x is None else str(x) diff --git a/TTS/utils/io.py b/TTS/utils/io.py deleted file mode 100644 index 3107ba661b..0000000000 --- a/TTS/utils/io.py +++ /dev/null @@ -1,70 +0,0 @@ -import os -import pickle as pickle_tts -from typing import Any, Callable, Dict, Union - -import fsspec -import torch - -from TTS.utils.generic_utils import get_user_data_dir - - -class RenamingUnpickler(pickle_tts.Unpickler): - """Overload default pickler to solve module renaming problem""" - - def find_class(self, module, name): - return super().find_class(module.replace("mozilla_voice_tts", "TTS"), name) - - -class AttrDict(dict): - """A custom dict which converts dict keys - to class attributes""" - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - self.__dict__ = self - - -def load_fsspec( - path: str, - map_location: Union[str, Callable, torch.device, Dict[Union[str, torch.device], Union[str, torch.device]]] = None, - cache: bool = True, - **kwargs, -) -> Any: - """Like torch.load but can load from other locations (e.g. s3:// , gs://). - - Args: - path: Any path or url supported by fsspec. - map_location: torch.device or str. - cache: If True, cache a remote file locally for subsequent calls. It is cached under `get_user_data_dir()/tts_cache`. Defaults to True. - **kwargs: Keyword arguments forwarded to torch.load. - - Returns: - Object stored in path. - """ - is_local = os.path.isdir(path) or os.path.isfile(path) - if cache and not is_local: - with fsspec.open( - f"filecache::{path}", - filecache={"cache_storage": str(get_user_data_dir("tts_cache"))}, - mode="rb", - ) as f: - return torch.load(f, map_location=map_location, **kwargs) - else: - with fsspec.open(path, "rb") as f: - return torch.load(f, map_location=map_location, **kwargs) - - -def load_checkpoint( - model, checkpoint_path, use_cuda=False, eval=False, cache=False -): # pylint: disable=redefined-builtin - try: - state = load_fsspec(checkpoint_path, map_location=torch.device("cpu"), cache=cache) - except ModuleNotFoundError: - pickle_tts.Unpickler = RenamingUnpickler - state = load_fsspec(checkpoint_path, map_location=torch.device("cpu"), pickle_module=pickle_tts, cache=cache) - model.load_state_dict(state["model"]) - if use_cuda: - model.cuda() - if eval: - model.eval() - return model, state diff --git a/TTS/utils/manage.py b/TTS/utils/manage.py index 3a527f4609..d7d4deab9d 100644 --- a/TTS/utils/manage.py +++ b/TTS/utils/manage.py @@ -1,18 +1,39 @@ import json +import logging import os import re import tarfile import zipfile from pathlib import Path from shutil import copyfile, rmtree -from typing import Dict, List, Tuple +from typing import Any, Optional, TypedDict, Union import fsspec import requests from tqdm import tqdm +from trainer.io import get_user_data_dir +from typing_extensions import Required from TTS.config import load_config, read_json_with_comments -from TTS.utils.generic_utils import get_user_data_dir + +logger = logging.getLogger(__name__) + + +class ModelItem(TypedDict, total=False): + model_name: Required[str] + model_type: Required[str] + description: str + license: str + author: str + contact: str + commit: Optional[str] + model_hash: str + tos_required: bool + default_vocoder: Optional[str] + model_url: Union[str, list[str]] + github_rls_url: Union[str, list[str]] + hf_url: list[str] + LICENSE_URLS = { "cc by-nc-nd 4.0": "https://creativecommons.org/licenses/by-nc-nd/4.0/", @@ -37,21 +58,24 @@ class ModelManager(object): home path. Args: - models_file (str): path to .model.json file. Defaults to None. - output_prefix (str): prefix to `tts` to download models. Defaults to None + models_file (str or Path): path to .model.json file. Defaults to None. + output_prefix (str or Path): prefix to `tts` to download models. Defaults to None progress_bar (bool): print a progress bar when donwloading a file. Defaults to False. - verbose (bool): print info. Defaults to True. """ - def __init__(self, models_file=None, output_prefix=None, progress_bar=False, verbose=True): + def __init__( + self, + models_file: Optional[Union[str, os.PathLike[Any]]] = None, + output_prefix: Optional[Union[str, os.PathLike[Any]]] = None, + progress_bar: bool = False, + ) -> None: super().__init__() self.progress_bar = progress_bar - self.verbose = verbose if output_prefix is None: self.output_prefix = get_user_data_dir("tts") else: - self.output_prefix = os.path.join(output_prefix, "tts") - self.models_dict = None + self.output_prefix = Path(output_prefix) / "tts" + self.models_dict = {} if models_file is not None: self.read_models_file(models_file) else: @@ -59,7 +83,7 @@ def __init__(self, models_file=None, output_prefix=None, progress_bar=False, ver path = Path(__file__).parent / "../.models.json" self.read_models_file(path) - def read_models_file(self, file_path): + def read_models_file(self, file_path: Union[str, os.PathLike[Any]]) -> None: """Read .models.json as a dict Args: @@ -67,53 +91,67 @@ def read_models_file(self, file_path): """ self.models_dict = read_json_with_comments(file_path) - def _list_models(self, model_type, model_count=0): - if self.verbose: - print("\n Name format: type/language/dataset/model") + def _list_models(self, model_type: str, model_count: int = 0) -> list[str]: + logger.info("") + logger.info("Name format: type/language/dataset/model") model_list = [] for lang in self.models_dict[model_type]: for dataset in self.models_dict[model_type][lang]: for model in self.models_dict[model_type][lang][dataset]: model_full_name = f"{model_type}--{lang}--{dataset}--{model}" - output_path = os.path.join(self.output_prefix, model_full_name) - if self.verbose: - if os.path.exists(output_path): - print(f" {model_count}: {model_type}/{lang}/{dataset}/{model} [already downloaded]") - else: - print(f" {model_count}: {model_type}/{lang}/{dataset}/{model}") + output_path = Path(self.output_prefix) / model_full_name + downloaded = " [already downloaded]" if output_path.is_dir() else "" + logger.info(" %2d: %s/%s/%s/%s%s", model_count, model_type, lang, dataset, model, downloaded) model_list.append(f"{model_type}/{lang}/{dataset}/{model}") model_count += 1 return model_list - def _list_for_model_type(self, model_type): + def _list_for_model_type(self, model_type: str) -> list[str]: models_name_list = [] model_count = 1 models_name_list.extend(self._list_models(model_type, model_count)) return models_name_list - def list_models(self): + def list_models(self) -> list[str]: models_name_list = [] model_count = 1 for model_type in self.models_dict: model_list = self._list_models(model_type, model_count) models_name_list.extend(model_list) + logger.info("") + logger.info("Path to downloaded models: %s", self.output_prefix) return models_name_list - def model_info_by_idx(self, model_query): - """Print the description of the model from .models.json file using model_idx + def log_model_details(self, model_type: str, lang: str, dataset: str, model: str) -> None: + logger.info("Model type: %s", model_type) + logger.info("Language supported: %s", lang) + logger.info("Dataset used: %s", dataset) + logger.info("Model name: %s", model) + if "description" in self.models_dict[model_type][lang][dataset][model]: + logger.info("Description: %s", self.models_dict[model_type][lang][dataset][model]["description"]) + else: + logger.info("Description: coming soon") + if "default_vocoder" in self.models_dict[model_type][lang][dataset][model]: + logger.info( + "Default vocoder: %s", + self.models_dict[model_type][lang][dataset][model]["default_vocoder"], + ) + + def model_info_by_idx(self, model_query: str) -> None: + """Print the description of the model from .models.json file using model_query_idx Args: - model_query (str): / + model_query (str): / """ model_name_list = [] model_type, model_query_idx = model_query.split("/") try: model_query_idx = int(model_query_idx) if model_query_idx <= 0: - print("> model_query_idx should be a positive integer!") + logger.error("model_query_idx [%d] should be a positive integer!", model_query_idx) return - except: - print("> model_query_idx should be an integer!") + except (TypeError, ValueError): + logger.error("model_query_idx [%s] should be an integer!", model_query_idx) return model_count = 0 if model_type in self.models_dict: @@ -123,153 +161,130 @@ def model_info_by_idx(self, model_query): model_name_list.append(f"{model_type}/{lang}/{dataset}/{model}") model_count += 1 else: - print(f"> model_type {model_type} does not exist in the list.") + logger.error("Model type %s does not exist in the list.", model_type) return if model_query_idx > model_count: - print(f"model query idx exceeds the number of available models [{model_count}] ") + logger.error("model_query_idx exceeds the number of available models [%d]", model_count) else: model_type, lang, dataset, model = model_name_list[model_query_idx - 1].split("/") - print(f"> model type : {model_type}") - print(f"> language supported : {lang}") - print(f"> dataset used : {dataset}") - print(f"> model name : {model}") - if "description" in self.models_dict[model_type][lang][dataset][model]: - print(f"> description : {self.models_dict[model_type][lang][dataset][model]['description']}") - else: - print("> description : coming soon") - if "default_vocoder" in self.models_dict[model_type][lang][dataset][model]: - print(f"> default_vocoder : {self.models_dict[model_type][lang][dataset][model]['default_vocoder']}") + self.log_model_details(model_type, lang, dataset, model) - def model_info_by_full_name(self, model_query_name): + def model_info_by_full_name(self, model_query_name: str) -> None: """Print the description of the model from .models.json file using model_full_name Args: model_query_name (str): Format is /// """ model_type, lang, dataset, model = model_query_name.split("/") - if model_type in self.models_dict: - if lang in self.models_dict[model_type]: - if dataset in self.models_dict[model_type][lang]: - if model in self.models_dict[model_type][lang][dataset]: - print(f"> model type : {model_type}") - print(f"> language supported : {lang}") - print(f"> dataset used : {dataset}") - print(f"> model name : {model}") - if "description" in self.models_dict[model_type][lang][dataset][model]: - print( - f"> description : {self.models_dict[model_type][lang][dataset][model]['description']}" - ) - else: - print("> description : coming soon") - if "default_vocoder" in self.models_dict[model_type][lang][dataset][model]: - print( - f"> default_vocoder : {self.models_dict[model_type][lang][dataset][model]['default_vocoder']}" - ) - else: - print(f"> model {model} does not exist for {model_type}/{lang}/{dataset}.") - else: - print(f"> dataset {dataset} does not exist for {model_type}/{lang}.") - else: - print(f"> lang {lang} does not exist for {model_type}.") - else: - print(f"> model_type {model_type} does not exist in the list.") + if model_type not in self.models_dict: + logger.error("Model type %s does not exist in the list.", model_type) + return + if lang not in self.models_dict[model_type]: + logger.error("Language %s does not exist for %s.", lang, model_type) + return + if dataset not in self.models_dict[model_type][lang]: + logger.error("Dataset %s does not exist for %s/%s.", dataset, model_type, lang) + return + if model not in self.models_dict[model_type][lang][dataset]: + logger.error("Model %s does not exist for %s/%s/%s.", model, model_type, lang, dataset) + return + self.log_model_details(model_type, lang, dataset, model) - def list_tts_models(self): + def list_tts_models(self) -> list[str]: """Print all `TTS` models and return a list of model names Format is `language/dataset/model` """ return self._list_for_model_type("tts_models") - def list_vocoder_models(self): + def list_vocoder_models(self) -> list[str]: """Print all the `vocoder` models and return a list of model names Format is `language/dataset/model` """ return self._list_for_model_type("vocoder_models") - def list_vc_models(self): + def list_vc_models(self) -> list[str]: """Print all the voice conversion models and return a list of model names Format is `language/dataset/model` """ return self._list_for_model_type("voice_conversion_models") - def list_langs(self): + def list_langs(self) -> None: """Print all the available languages""" - print(" Name format: type/language") + logger.info("Name format: type/language") for model_type in self.models_dict: for lang in self.models_dict[model_type]: - print(f" >: {model_type}/{lang} ") + logger.info(" %s/%s", model_type, lang) - def list_datasets(self): + def list_datasets(self) -> None: """Print all the datasets""" - print(" Name format: type/language/dataset") + logger.info("Name format: type/language/dataset") for model_type in self.models_dict: for lang in self.models_dict[model_type]: for dataset in self.models_dict[model_type][lang]: - print(f" >: {model_type}/{lang}/{dataset}") + logger.info(" %s/%s/%s", model_type, lang, dataset) @staticmethod - def print_model_license(model_item: Dict): + def print_model_license(model_item: ModelItem) -> None: """Print the license of a model Args: model_item (dict): model item in the models.json """ if "license" in model_item and model_item["license"].strip() != "": - print(f" > Model's license - {model_item['license']}") + logger.info("Model's license - %s", model_item["license"]) if model_item["license"].lower() in LICENSE_URLS: - print(f" > Check {LICENSE_URLS[model_item['license'].lower()]} for more info.") + logger.info("Check %s for more info.", LICENSE_URLS[model_item["license"].lower()]) else: - print(" > Check https://opensource.org/licenses for more info.") + logger.info("Check https://opensource.org/licenses for more info.") else: - print(" > Model's license - No license information available") + logger.info("Model's license - No license information available") - def _download_github_model(self, model_item: Dict, output_path: str): + def _download_github_model(self, model_item: ModelItem, output_path: Path) -> None: if isinstance(model_item["github_rls_url"], list): self._download_model_files(model_item["github_rls_url"], output_path, self.progress_bar) else: self._download_zip_file(model_item["github_rls_url"], output_path, self.progress_bar) - def _download_hf_model(self, model_item: Dict, output_path: str): + def _download_hf_model(self, model_item: ModelItem, output_path: Path) -> None: if isinstance(model_item["hf_url"], list): self._download_model_files(model_item["hf_url"], output_path, self.progress_bar) else: self._download_zip_file(model_item["hf_url"], output_path, self.progress_bar) - def download_fairseq_model(self, model_name, output_path): - URI_PREFIX = "https://coqui.gateway.scarf.sh/fairseq/" + def download_fairseq_model(self, model_name: str, output_path: Path) -> None: + URI_PREFIX = "https://dl.fbaipublicfiles.com/mms/tts/" _, lang, _, _ = model_name.split("/") model_download_uri = os.path.join(URI_PREFIX, f"{lang}.tar.gz") self._download_tar_file(model_download_uri, output_path, self.progress_bar) @staticmethod - def set_model_url(model_item: Dict): - model_item["model_url"] = None + def set_model_url(model_item: ModelItem) -> ModelItem: + model_item["model_url"] = "" if "github_rls_url" in model_item: model_item["model_url"] = model_item["github_rls_url"] elif "hf_url" in model_item: model_item["model_url"] = model_item["hf_url"] elif "fairseq" in model_item["model_name"]: - model_item["model_url"] = "https://coqui.gateway.scarf.sh/fairseq/" + model_item["model_url"] = "https://dl.fbaipublicfiles.com/mms/tts/" elif "xtts" in model_item["model_name"]: - model_item["model_url"] = "https://coqui.gateway.scarf.sh/xtts/" + model_item["model_url"] = "https://huggingface.co/coqui/" return model_item - def _set_model_item(self, model_name): + def _set_model_item(self, model_name: str) -> tuple[ModelItem, str, str, Optional[str]]: # fetch model info from the dict if "fairseq" in model_name: - model_type = "tts_models" - lang = model_name.split("/")[1] - model_item = { + model_type, lang, dataset, model = model_name.split("/") + model_item: ModelItem = { + "model_name": model_name, "model_type": "tts_models", "license": "CC BY-NC 4.0", "default_vocoder": None, "author": "fairseq", "description": "this model is released by Meta under Fairseq repo. Visit https://github.com/facebookresearch/fairseq/tree/main/examples/mms for more info.", } - model_item["model_name"] = model_name elif "xtts" in model_name and len(model_name.split("/")) != 4: # loading xtts models with only model name (e.g. xtts_v2.0.2) # check model name has the version number with regex @@ -283,16 +298,18 @@ def _set_model_item(self, model_name): dataset = "multi-dataset" model = model_name model_item = { + "model_name": model_name, + "model_type": model_type, "default_vocoder": None, "license": "CPML", "contact": "info@coqui.ai", "tos_required": True, "hf_url": [ - f"https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/{model_version}/model.pth", - f"https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/{model_version}/config.json", - f"https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/{model_version}/vocab.json", - f"https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/{model_version}/hash.md5", - f"https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/{model_version}/speakers_xtts.pth", + f"https://huggingface.co/coqui/XTTS-v2/resolve/{model_version}/model.pth", + f"https://huggingface.co/coqui/XTTS-v2/resolve/{model_version}/config.json", + f"https://huggingface.co/coqui/XTTS-v2/resolve/{model_version}/vocab.json", + f"https://huggingface.co/coqui/XTTS-v2/resolve/{model_version}/hash.md5", + f"https://huggingface.co/coqui/XTTS-v2/resolve/{model_version}/speakers_xtts.pth", ], } else: @@ -307,9 +324,9 @@ def _set_model_item(self, model_name): return model_item, model_full_name, model, md5hash @staticmethod - def ask_tos(model_full_path): + def ask_tos(model_full_path: Path) -> bool: """Ask the user to agree to the terms of service""" - tos_path = os.path.join(model_full_path, "tos_agreed.txt") + tos_path = model_full_path / "tos_agreed.txt" print(" > You must confirm the following:") print(' | > "I have purchased a commercial license from Coqui: licensing@coqui.ai"') print(' | > "Otherwise, I agree to the terms of the non-commercial CPML: https://coqui.ai/cpml" - [y/n]') @@ -321,7 +338,7 @@ def ask_tos(model_full_path): return False @staticmethod - def tos_agreed(model_item, model_full_path): + def tos_agreed(model_item: ModelItem, model_full_path: Path) -> bool: """Check if the user has agreed to the terms of service""" if "tos_required" in model_item and model_item["tos_required"]: tos_path = os.path.join(model_full_path, "tos_agreed.txt") @@ -330,14 +347,14 @@ def tos_agreed(model_item, model_full_path): return False return True - def create_dir_and_download_model(self, model_name, model_item, output_path): - os.makedirs(output_path, exist_ok=True) + def create_dir_and_download_model(self, model_name: str, model_item: ModelItem, output_path: Path) -> None: + output_path.mkdir(exist_ok=True, parents=True) # handle TOS if not self.tos_agreed(model_item, output_path): if not self.ask_tos(output_path): - os.rmdir(output_path) + output_path.rmdir() raise Exception(" [!] You must agree to the terms of service to use this model.") - print(f" > Downloading model to {output_path}") + logger.info("Downloading model to %s", output_path) try: if "fairseq" in model_name: self.download_fairseq_model(model_name, output_path) @@ -347,12 +364,12 @@ def create_dir_and_download_model(self, model_name, model_item, output_path): self._download_hf_model(model_item, output_path) except requests.RequestException as e: - print(f" > Failed to download the model file to {output_path}") + logger.exception("Failed to download the model file to %s", output_path) rmtree(output_path) raise e self.print_model_license(model_item=model_item) - def check_if_configs_are_equal(self, model_name, model_item, output_path): + def check_if_configs_are_equal(self, model_name: str, model_item: ModelItem, output_path: Path) -> None: with fsspec.open(self._find_files(output_path)[1], "r", encoding="utf-8") as f: config_local = json.load(f) remote_url = None @@ -365,10 +382,10 @@ def check_if_configs_are_equal(self, model_name, model_item, output_path): config_remote = json.load(f) if not config_local == config_remote: - print(f" > {model_name} is already downloaded however it has been changed. Redownloading it...") + logger.info("%s is already downloaded however it has been changed. Redownloading it...", model_name) self.create_dir_and_download_model(model_name, model_item, output_path) - def download_model(self, model_name): + def download_model(self, model_name: str) -> tuple[Path, Optional[Path], ModelItem]: """Download model files given the full model name. Model name is in the format 'type/language/dataset/model' @@ -384,19 +401,19 @@ def download_model(self, model_name): """ model_item, model_full_name, model, md5sum = self._set_model_item(model_name) # set the model specific output path - output_path = os.path.join(self.output_prefix, model_full_name) - if os.path.exists(output_path): + output_path = Path(self.output_prefix) / model_full_name + if output_path.is_dir(): if md5sum is not None: - md5sum_file = os.path.join(output_path, "hash.md5") - if os.path.isfile(md5sum_file): - with open(md5sum_file, mode="r") as f: + md5sum_file = output_path / "hash.md5" + if md5sum_file.is_file(): + with md5sum_file.open() as f: if not f.read() == md5sum: - print(f" > {model_name} has been updated, clearing model cache...") + logger.info("%s has been updated, clearing model cache...", model_name) self.create_dir_and_download_model(model_name, model_item, output_path) else: - print(f" > {model_name} is already downloaded.") + logger.info("%s is already downloaded.", model_name) else: - print(f" > {model_name} has been updated, clearing model cache...") + logger.info("%s has been updated, clearing model cache...", model_name) self.create_dir_and_download_model(model_name, model_item, output_path) # if the configs are different, redownload it # ToDo: we need a better way to handle it @@ -406,7 +423,7 @@ def download_model(self, model_name): except: pass else: - print(f" > {model_name} is already downloaded.") + logger.info("%s is already downloaded.", model_name) else: self.create_dir_and_download_model(model_name, model_item, output_path) @@ -417,12 +434,14 @@ def download_model(self, model_name): model not in ["tortoise-v2", "bark"] and "fairseq" not in model_name and "xtts" not in model_name ): # TODO:This is stupid but don't care for now. output_model_path, output_config_path = self._find_files(output_path) + else: + output_config_path = output_model_path / "config.json" # update paths in the config.json self._update_paths(output_path, output_config_path) return output_model_path, output_config_path, model_item @staticmethod - def _find_files(output_path: str) -> Tuple[str, str]: + def _find_files(output_path: Path) -> tuple[Path, Path]: """Find the model and config files in the output path Args: @@ -433,11 +452,11 @@ def _find_files(output_path: str) -> Tuple[str, str]: """ model_file = None config_file = None - for file_name in os.listdir(output_path): - if file_name in ["model_file.pth", "model_file.pth.tar", "model.pth"]: - model_file = os.path.join(output_path, file_name) - elif file_name == "config.json": - config_file = os.path.join(output_path, file_name) + for f in output_path.iterdir(): + if f.name in ["model_file.pth", "model_file.pth.tar", "model.pth", "checkpoint.pth"]: + model_file = f + elif f.name == "config.json": + config_file = f if model_file is None: raise ValueError(" [!] Model file not found in the output path") if config_file is None: @@ -445,7 +464,7 @@ def _find_files(output_path: str) -> Tuple[str, str]: return model_file, config_file @staticmethod - def _find_speaker_encoder(output_path: str) -> str: + def _find_speaker_encoder(output_path: Path) -> Optional[Path]: """Find the speaker encoder file in the output path Args: @@ -455,24 +474,24 @@ def _find_speaker_encoder(output_path: str) -> str: str: path to the speaker encoder file """ speaker_encoder_file = None - for file_name in os.listdir(output_path): - if file_name in ["model_se.pth", "model_se.pth.tar"]: - speaker_encoder_file = os.path.join(output_path, file_name) + for f in output_path.iterdir(): + if f.name in ["model_se.pth", "model_se.pth.tar"]: + speaker_encoder_file = f return speaker_encoder_file - def _update_paths(self, output_path: str, config_path: str) -> None: + def _update_paths(self, output_path: Path, config_path: Path) -> None: """Update paths for certain files in config.json after download. Args: output_path (str): local path the model is downloaded to. config_path (str): local config.json path. """ - output_stats_path = os.path.join(output_path, "scale_stats.npy") - output_d_vector_file_path = os.path.join(output_path, "speakers.json") - output_d_vector_file_pth_path = os.path.join(output_path, "speakers.pth") - output_speaker_ids_file_path = os.path.join(output_path, "speaker_ids.json") - output_speaker_ids_file_pth_path = os.path.join(output_path, "speaker_ids.pth") - speaker_encoder_config_path = os.path.join(output_path, "config_se.json") + output_stats_path = output_path / "scale_stats.npy" + output_d_vector_file_path = output_path / "speakers.json" + output_d_vector_file_pth_path = output_path / "speakers.pth" + output_speaker_ids_file_path = output_path / "speaker_ids.json" + output_speaker_ids_file_pth_path = output_path / "speaker_ids.pth" + speaker_encoder_config_path = output_path / "config_se.json" speaker_encoder_model_path = self._find_speaker_encoder(output_path) # update the scale_path.npy file path in the model config.json @@ -497,10 +516,10 @@ def _update_paths(self, output_path: str, config_path: str) -> None: self._update_path("model_args.speaker_encoder_config_path", speaker_encoder_config_path, config_path) @staticmethod - def _update_path(field_name, new_path, config_path): + def _update_path(field_name: str, new_path: Optional[Path], config_path: Path) -> None: """Update the path in the model config.json for the current environment after download""" - if new_path and os.path.exists(new_path): - config = load_config(config_path) + if new_path is not None and new_path.is_file(): + config = load_config(str(config_path)) field_names = field_name.split(".") if len(field_names) > 1: # field name points to a sub-level field @@ -516,7 +535,7 @@ def _update_path(field_name, new_path, config_path): sub_conf[field_names[-1]] = new_path else: # field name points to a top-level field - if not field_name in config: + if field_name not in config: return if isinstance(config[field_name], list): config[field_name] = [new_path] @@ -525,7 +544,7 @@ def _update_path(field_name, new_path, config_path): config.save_json(config_path) @staticmethod - def _download_zip_file(file_url, output_folder, progress_bar): + def _download_zip_file(file_url: str, output_folder: Path, progress_bar: bool) -> None: """Download the github releases""" # download the file r = requests.get(file_url, stream=True) @@ -535,7 +554,7 @@ def _download_zip_file(file_url, output_folder, progress_bar): block_size = 1024 # 1 Kibibyte if progress_bar: ModelManager.tqdm_progress = tqdm(total=total_size_in_bytes, unit="iB", unit_scale=True) - temp_zip_name = os.path.join(output_folder, file_url.split("/")[-1]) + temp_zip_name = output_folder / file_url.split("/")[-1] with open(temp_zip_name, "wb") as file: for data in r.iter_content(block_size): if progress_bar: @@ -543,24 +562,24 @@ def _download_zip_file(file_url, output_folder, progress_bar): file.write(data) with zipfile.ZipFile(temp_zip_name) as z: z.extractall(output_folder) - os.remove(temp_zip_name) # delete zip after extract + temp_zip_name.unlink() # delete zip after extract except zipfile.BadZipFile: - print(f" > Error: Bad zip file - {file_url}") + logger.exception("Bad zip file - %s", file_url) raise zipfile.BadZipFile # pylint: disable=raise-missing-from # move the files to the outer path for file_path in z.namelist(): - src_path = os.path.join(output_folder, file_path) - if os.path.isfile(src_path): - dst_path = os.path.join(output_folder, os.path.basename(file_path)) + src_path = output_folder / file_path + if src_path.is_file(): + dst_path = output_folder / os.path.basename(file_path) if src_path != dst_path: copyfile(src_path, dst_path) # remove redundant (hidden or not) folders for file_path in z.namelist(): - if os.path.isdir(os.path.join(output_folder, file_path)): - rmtree(os.path.join(output_folder, file_path)) + if (output_folder / file_path).is_dir(): + rmtree(output_folder / file_path) @staticmethod - def _download_tar_file(file_url, output_folder, progress_bar): + def _download_tar_file(file_url: str, output_folder: Path, progress_bar: bool) -> None: """Download the github releases""" # download the file r = requests.get(file_url, stream=True) @@ -570,7 +589,7 @@ def _download_tar_file(file_url, output_folder, progress_bar): block_size = 1024 # 1 Kibibyte if progress_bar: ModelManager.tqdm_progress = tqdm(total=total_size_in_bytes, unit="iB", unit_scale=True) - temp_tar_name = os.path.join(output_folder, file_url.split("/")[-1]) + temp_tar_name = output_folder / file_url.split("/")[-1] with open(temp_tar_name, "wb") as file: for data in r.iter_content(block_size): if progress_bar: @@ -579,43 +598,37 @@ def _download_tar_file(file_url, output_folder, progress_bar): with tarfile.open(temp_tar_name) as t: t.extractall(output_folder) tar_names = t.getnames() - os.remove(temp_tar_name) # delete tar after extract + temp_tar_name.unlink() # delete tar after extract except tarfile.ReadError: - print(f" > Error: Bad tar file - {file_url}") + logger.exception("Bad tar file - %s", file_url) raise tarfile.ReadError # pylint: disable=raise-missing-from # move the files to the outer path - for file_path in os.listdir(os.path.join(output_folder, tar_names[0])): - src_path = os.path.join(output_folder, tar_names[0], file_path) - dst_path = os.path.join(output_folder, os.path.basename(file_path)) + for file_path in (output_folder / tar_names[0]).iterdir(): + src_path = file_path + dst_path = output_folder / file_path.name if src_path != dst_path: copyfile(src_path, dst_path) # remove the extracted folder - rmtree(os.path.join(output_folder, tar_names[0])) + rmtree(output_folder / tar_names[0]) @staticmethod - def _download_model_files(file_urls, output_folder, progress_bar): + def _download_model_files( + file_urls: list[str], output_folder: Union[str, os.PathLike[Any]], progress_bar: bool + ) -> None: """Download the github releases""" + output_folder = Path(output_folder) for file_url in file_urls: # download the file r = requests.get(file_url, stream=True) # extract the file - bease_filename = file_url.split("/")[-1] - temp_zip_name = os.path.join(output_folder, bease_filename) + base_filename = file_url.split("/")[-1] + file_path = output_folder / base_filename total_size_in_bytes = int(r.headers.get("content-length", 0)) block_size = 1024 # 1 Kibibyte - with open(temp_zip_name, "wb") as file: + with open(file_path, "wb") as f: if progress_bar: ModelManager.tqdm_progress = tqdm(total=total_size_in_bytes, unit="iB", unit_scale=True) for data in r.iter_content(block_size): if progress_bar: ModelManager.tqdm_progress.update(len(data)) - file.write(data) - - @staticmethod - def _check_dict_key(my_dict, key): - if key in my_dict.keys() and my_dict[key] is not None: - if not isinstance(key, str): - return True - if isinstance(key, str) and len(my_dict[key]) > 0: - return True - return False + f.write(data) diff --git a/TTS/utils/synthesizer.py b/TTS/utils/synthesizer.py index b98647c30c..517cb7d2b2 100644 --- a/TTS/utils/synthesizer.py +++ b/TTS/utils/synthesizer.py @@ -1,6 +1,8 @@ +import logging import os import time -from typing import List +from pathlib import Path +from typing import Any, List, Optional, Union import numpy as np import pysbd @@ -11,32 +13,35 @@ from TTS.tts.configs.vits_config import VitsConfig from TTS.tts.models import setup_model as setup_tts_model from TTS.tts.models.vits import Vits - -# pylint: disable=unused-wildcard-import -# pylint: disable=wildcard-import from TTS.tts.utils.synthesis import synthesis, transfer_voice, trim_silence from TTS.utils.audio import AudioProcessor from TTS.utils.audio.numpy_transforms import save_wav +from TTS.utils.generic_utils import optional_to_str +from TTS.vc.configs.openvoice_config import OpenVoiceConfig from TTS.vc.models import setup_model as setup_vc_model +from TTS.vc.models.openvoice import OpenVoice from TTS.vocoder.models import setup_model as setup_vocoder_model from TTS.vocoder.utils.generic_utils import interpolate_vocoder_input +logger = logging.getLogger(__name__) + class Synthesizer(nn.Module): def __init__( self, - tts_checkpoint: str = "", - tts_config_path: str = "", - tts_speakers_file: str = "", - tts_languages_file: str = "", - vocoder_checkpoint: str = "", - vocoder_config: str = "", - encoder_checkpoint: str = "", - encoder_config: str = "", - vc_checkpoint: str = "", - vc_config: str = "", - model_dir: str = "", - voice_dir: str = None, + *, + tts_checkpoint: Optional[Union[str, os.PathLike[Any]]] = None, + tts_config_path: Optional[Union[str, os.PathLike[Any]]] = None, + tts_speakers_file: Optional[Union[str, os.PathLike[Any]]] = None, + tts_languages_file: Optional[Union[str, os.PathLike[Any]]] = None, + vocoder_checkpoint: Optional[Union[str, os.PathLike[Any]]] = None, + vocoder_config: Optional[Union[str, os.PathLike[Any]]] = None, + encoder_checkpoint: Optional[Union[str, os.PathLike[Any]]] = None, + encoder_config: Optional[Union[str, os.PathLike[Any]]] = None, + vc_checkpoint: Optional[Union[str, os.PathLike[Any]]] = None, + vc_config: Optional[Union[str, os.PathLike[Any]]] = None, + model_dir: Optional[Union[str, os.PathLike[Any]]] = None, + voice_dir: Optional[Union[str, os.PathLike[Any]]] = None, use_cuda: bool = False, ) -> None: """General 🐸 TTS interface for inference. It takes a tts and a vocoder @@ -62,16 +67,17 @@ def __init__( use_cuda (bool, optional): enable/disable cuda. Defaults to False. """ super().__init__() - self.tts_checkpoint = tts_checkpoint - self.tts_config_path = tts_config_path - self.tts_speakers_file = tts_speakers_file - self.tts_languages_file = tts_languages_file - self.vocoder_checkpoint = vocoder_checkpoint - self.vocoder_config = vocoder_config - self.encoder_checkpoint = encoder_checkpoint - self.encoder_config = encoder_config - self.vc_checkpoint = vc_checkpoint - self.vc_config = vc_config + self.tts_checkpoint = optional_to_str(tts_checkpoint) + self.tts_config_path = optional_to_str(tts_config_path) + self.tts_speakers_file = optional_to_str(tts_speakers_file) + self.tts_languages_file = optional_to_str(tts_languages_file) + self.vocoder_checkpoint = optional_to_str(vocoder_checkpoint) + self.vocoder_config = optional_to_str(vocoder_config) + self.encoder_checkpoint = optional_to_str(encoder_checkpoint) + self.encoder_config = optional_to_str(encoder_config) + self.vc_checkpoint = optional_to_str(vc_checkpoint) + self.vc_config = optional_to_str(vc_config) + model_dir = optional_to_str(model_dir) self.use_cuda = use_cuda self.tts_model = None @@ -90,24 +96,21 @@ def __init__( assert torch.cuda.is_available(), "CUDA is not availabe on this machine." if tts_checkpoint: - self._load_tts(tts_checkpoint, tts_config_path, use_cuda) - self.output_sample_rate = self.tts_config.audio["sample_rate"] + self._load_tts(self.tts_checkpoint, self.tts_config_path, use_cuda) if vocoder_checkpoint: - self._load_vocoder(vocoder_checkpoint, vocoder_config, use_cuda) - self.output_sample_rate = self.vocoder_config.audio["sample_rate"] + self._load_vocoder(self.vocoder_checkpoint, self.vocoder_config, use_cuda) - if vc_checkpoint: - self._load_vc(vc_checkpoint, vc_config, use_cuda) - self.output_sample_rate = self.vc_config.audio["output_sample_rate"] + if vc_checkpoint and model_dir == "": + self._load_vc(self.vc_checkpoint, self.vc_config, use_cuda) if model_dir: if "fairseq" in model_dir: self._load_fairseq_from_dir(model_dir, use_cuda) - self.output_sample_rate = self.tts_config.audio["sample_rate"] + elif "openvoice" in model_dir: + self._load_openvoice_from_dir(Path(model_dir), use_cuda) else: self._load_tts_from_dir(model_dir, use_cuda) - self.output_sample_rate = self.tts_config.audio["output_sample_rate"] @staticmethod def _get_segmenter(lang: str): @@ -136,6 +139,7 @@ def _load_vc(self, vc_checkpoint: str, vc_config_path: str, use_cuda: bool) -> N """ # pylint: disable=global-statement self.vc_config = load_config(vc_config_path) + self.output_sample_rate = self.vc_config.audio["output_sample_rate"] self.vc_model = setup_vc_model(config=self.vc_config) self.vc_model.load_checkpoint(self.vc_config, vc_checkpoint) if use_cuda: @@ -150,9 +154,24 @@ def _load_fairseq_from_dir(self, model_dir: str, use_cuda: bool) -> None: self.tts_model = Vits.init_from_config(self.tts_config) self.tts_model.load_fairseq_checkpoint(self.tts_config, checkpoint_dir=model_dir, eval=True) self.tts_config = self.tts_model.config + self.output_sample_rate = self.tts_config.audio["sample_rate"] if use_cuda: self.tts_model.cuda() + def _load_openvoice_from_dir(self, checkpoint: Path, use_cuda: bool) -> None: + """Load the OpenVoice model from a directory. + + We assume the model knows how to load itself from the directory and + there is a config.json file in the directory. + """ + self.vc_config = OpenVoiceConfig() + self.vc_model = OpenVoice.init_from_config(self.vc_config) + self.vc_model.load_checkpoint(self.vc_config, checkpoint, eval=True) + self.vc_config = self.vc_model.config + self.output_sample_rate = self.vc_config.audio["output_sample_rate"] + if use_cuda: + self.vc_model.cuda() + def _load_tts_from_dir(self, model_dir: str, use_cuda: bool) -> None: """Load the TTS model from a directory. @@ -160,6 +179,7 @@ def _load_tts_from_dir(self, model_dir: str, use_cuda: bool) -> None: """ config = load_config(os.path.join(model_dir, "config.json")) self.tts_config = config + self.output_sample_rate = self.tts_config.audio["output_sample_rate"] self.tts_model = setup_tts_model(config) self.tts_model.load_checkpoint(config, checkpoint_dir=model_dir, eval=True) if use_cuda: @@ -181,6 +201,7 @@ def _load_tts(self, tts_checkpoint: str, tts_config_path: str, use_cuda: bool) - """ # pylint: disable=global-statement self.tts_config = load_config(tts_config_path) + self.output_sample_rate = self.tts_config.audio["sample_rate"] if self.tts_config["use_phonemes"] and self.tts_config["phonemizer"] is None: raise ValueError("Phonemizer is not defined in the TTS config.") @@ -218,7 +239,8 @@ def _load_vocoder(self, model_file: str, model_config: str, use_cuda: bool) -> N use_cuda (bool): enable/disable CUDA use. """ self.vocoder_config = load_config(model_config) - self.vocoder_ap = AudioProcessor(verbose=False, **self.vocoder_config.audio) + self.output_sample_rate = self.vocoder_config.audio["sample_rate"] + self.vocoder_ap = AudioProcessor(**self.vocoder_config.audio) self.vocoder_model = setup_vocoder_model(self.vocoder_config) self.vocoder_model.load_checkpoint(self.vocoder_config, model_file, eval=True) if use_cuda: @@ -294,9 +316,9 @@ def tts( if text: sens = [text] if split_sentences: - print(" > Text splitted to sentences.") sens = self.split_into_sentences(text) - print(sens) + logger.info("Text split into sentences.") + logger.info("Input: %s", sens) # handle multi-speaker if "voice_dir" in kwargs: @@ -335,7 +357,7 @@ def tts( # handle multi-lingual language_id = None if self.tts_languages_file or ( - hasattr(self.tts_model, "language_manager") + hasattr(self.tts_model, "language_manager") and self.tts_model.language_manager is not None and not self.tts_config.model == "xtts" ): @@ -420,7 +442,7 @@ def tts( self.vocoder_config["audio"]["sample_rate"] / self.tts_model.ap.sample_rate, ] if scale_factor[1] != 1: - print(" > interpolating tts model output.") + logger.info("Interpolating TTS model output.") vocoder_input = interpolate_vocoder_input(scale_factor, vocoder_input) else: vocoder_input = torch.tensor(vocoder_input).unsqueeze(0) # pylint: disable=not-callable @@ -484,7 +506,7 @@ def tts( self.vocoder_config["audio"]["sample_rate"] / self.tts_model.ap.sample_rate, ] if scale_factor[1] != 1: - print(" > interpolating tts model output.") + logger.info("Interpolating TTS model output.") vocoder_input = interpolate_vocoder_input(scale_factor, vocoder_input) else: vocoder_input = torch.tensor(vocoder_input).unsqueeze(0) # pylint: disable=not-callable @@ -500,6 +522,6 @@ def tts( # compute stats process_time = time.time() - start_time audio_time = len(wavs) / self.tts_config.audio["sample_rate"] - print(f" > Processing time: {process_time}") - print(f" > Real-time factor: {process_time / audio_time}") + logger.info("Processing time: %.3f", process_time) + logger.info("Real-time factor: %.3f", process_time / audio_time) return wavs diff --git a/TTS/utils/training.py b/TTS/utils/training.py index b51f55e92b..57885005f1 100644 --- a/TTS/utils/training.py +++ b/TTS/utils/training.py @@ -1,6 +1,10 @@ +import logging + import numpy as np import torch +logger = logging.getLogger(__name__) + def check_update(model, grad_clip, ignore_stopnet=False, amp_opt_params=None): r"""Check model gradient against unexpected jumps and failures""" @@ -21,11 +25,11 @@ def check_update(model, grad_clip, ignore_stopnet=False, amp_opt_params=None): # compatibility with different torch versions if isinstance(grad_norm, float): if np.isinf(grad_norm): - print(" | > Gradient is INF !!") + logger.warning("Gradient is INF !!") skip_flag = True else: if torch.isinf(grad_norm): - print(" | > Gradient is INF !!") + logger.warning("Gradient is INF !!") skip_flag = True return grad_norm, skip_flag diff --git a/TTS/utils/vad.py b/TTS/utils/vad.py index aefce2b50b..49c8dc6b66 100644 --- a/TTS/utils/vad.py +++ b/TTS/utils/vad.py @@ -1,6 +1,10 @@ +import logging + import torch import torchaudio +logger = logging.getLogger(__name__) + def read_audio(path): wav, sr = torchaudio.load(path) @@ -54,8 +58,8 @@ def remove_silence( # read ground truth wav and resample the audio for the VAD try: wav, gt_sample_rate = read_audio(audio_path) - except: - print(f"> ❗ Failed to read {audio_path}") + except Exception: + logger.exception("Failed to read %s", audio_path) return None, False # if needed, resample the audio for the VAD model @@ -80,7 +84,7 @@ def remove_silence( wav = collect_chunks(new_speech_timestamps, wav) is_speech = True else: - print(f"> The file {audio_path} probably does not have speech please check it !!") + logger.warning("The file %s probably does not have speech please check it!", audio_path) is_speech = False # save diff --git a/TTS/vc/configs/freevc_config.py b/TTS/vc/configs/freevc_config.py index 207181b303..d600bfb1f4 100644 --- a/TTS/vc/configs/freevc_config.py +++ b/TTS/vc/configs/freevc_config.py @@ -229,7 +229,7 @@ class FreeVCConfig(BaseVCConfig): If true, language embedding is used. Defaults to `False`. Note: - Check :class:`TTS.tts.configs.shared_configs.BaseTTSConfig` for the inherited parameters. + Check :class:`TTS.tts.configs.shared_configs.BaseVCConfig` for the inherited parameters. Example: diff --git a/TTS/vc/configs/openvoice_config.py b/TTS/vc/configs/openvoice_config.py new file mode 100644 index 0000000000..261cdd6f47 --- /dev/null +++ b/TTS/vc/configs/openvoice_config.py @@ -0,0 +1,201 @@ +from dataclasses import dataclass, field +from typing import Optional + +from coqpit import Coqpit + +from TTS.vc.configs.shared_configs import BaseVCConfig + + +@dataclass +class OpenVoiceAudioConfig(Coqpit): + """Audio configuration + + Args: + input_sample_rate (int): + The sampling rate of the input waveform. + + output_sample_rate (int): + The sampling rate of the output waveform. + + fft_size (int): + The length of the filter. + + hop_length (int): + The hop length. + + win_length (int): + The window length. + """ + + input_sample_rate: int = field(default=22050) + output_sample_rate: int = field(default=22050) + fft_size: int = field(default=1024) + hop_length: int = field(default=256) + win_length: int = field(default=1024) + + +@dataclass +class OpenVoiceArgs(Coqpit): + """OpenVoice model arguments. + + zero_g (bool): + Whether to zero the gradients. + + inter_channels (int): + The number of channels in the intermediate layers. + + hidden_channels (int): + The number of channels in the hidden layers. + + filter_channels (int): + The number of channels in the filter layers. + + n_heads (int): + The number of attention heads. + + n_layers (int): + The number of layers. + + kernel_size (int): + The size of the kernel. + + p_dropout (float): + The dropout probability. + + resblock (str): + The type of residual block. + + resblock_kernel_sizes (List[int]): + The kernel sizes for the residual blocks. + + resblock_dilation_sizes (List[List[int]]): + The dilation sizes for the residual blocks. + + upsample_rates (List[int]): + The upsample rates. + + upsample_initial_channel (int): + The number of channels in the initial upsample layer. + + upsample_kernel_sizes (List[int]): + The kernel sizes for the upsample layers. + + n_layers_q (int): + The number of layers in the quantization network. + + use_spectral_norm (bool): + Whether to use spectral normalization. + + gin_channels (int): + The number of channels in the global conditioning vector. + + tau (float): + Tau parameter for the posterior encoder + """ + + zero_g: bool = field(default=True) + inter_channels: int = field(default=192) + hidden_channels: int = field(default=192) + filter_channels: int = field(default=768) + n_heads: int = field(default=2) + n_layers: int = field(default=6) + kernel_size: int = field(default=3) + p_dropout: float = field(default=0.1) + resblock: str = field(default="1") + resblock_kernel_sizes: list[int] = field(default_factory=lambda: [3, 7, 11]) + resblock_dilation_sizes: list[list[int]] = field(default_factory=lambda: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]) + upsample_rates: list[int] = field(default_factory=lambda: [8, 8, 2, 2]) + upsample_initial_channel: int = field(default=512) + upsample_kernel_sizes: list[int] = field(default_factory=lambda: [16, 16, 4, 4]) + n_layers_q: int = field(default=3) + use_spectral_norm: bool = field(default=False) + gin_channels: int = field(default=256) + tau: float = field(default=0.3) + + +@dataclass +class OpenVoiceConfig(BaseVCConfig): + """Defines parameters for OpenVoice VC model. + + Args: + model (str): + Model name. Do not change unless you know what you are doing. + + model_args (OpenVoiceArgs): + Model architecture arguments. Defaults to `OpenVoiceArgs()`. + + audio (OpenVoiceAudioConfig): + Audio processing configuration. Defaults to `OpenVoiceAudioConfig()`. + + return_wav (bool): + If true, data loader returns the waveform as well as the other outputs. Do not change. Defaults to `True`. + + compute_linear_spec (bool): + If true, the linear spectrogram is computed and returned alongside the mel output. Do not change. Defaults to `True`. + + use_weighted_sampler (bool): + If true, use weighted sampler with bucketing for balancing samples between datasets used in training. Defaults to `False`. + + weighted_sampler_attrs (dict): + Key retuned by the formatter to be used for weighted sampler. For example `{"root_path": 2.0, "speaker_name": 1.0}` sets sample probabilities + by overweighting `root_path` by 2.0. Defaults to `{}`. + + weighted_sampler_multipliers (dict): + Weight each unique value of a key returned by the formatter for weighted sampling. + For example `{"root_path":{"/raid/datasets/libritts-clean-16khz-bwe-coqui_44khz/LibriTTS/train-clean-100/":1.0, "/raid/datasets/libritts-clean-16khz-bwe-coqui_44khz/LibriTTS/train-clean-360/": 0.5}`. + It will sample instances from `train-clean-100` 2 times more than `train-clean-360`. Defaults to `{}`. + + r (int): + Number of spectrogram frames to be generated at a time. Do not change. Defaults to `1`. + + add_blank (bool): + If true, a blank token is added in between every character. Defaults to `True`. + + Note: + Check :class:`TTS.tts.configs.shared_configs.BaseVCConfig` for the inherited parameters. + + Example: + + >>> from TTS.vc.configs.openvoice_config import OpenVoiceConfig + >>> config = OpenVoiceConfig() + """ + + model: str = "openvoice" + # model specific params + model_args: OpenVoiceArgs = field(default_factory=OpenVoiceArgs) + audio: OpenVoiceAudioConfig = field(default_factory=OpenVoiceAudioConfig) + + # optimizer + # TODO with training support + + # loss params + # TODO with training support + + # data loader params + return_wav: bool = True + compute_linear_spec: bool = True + + # sampler params + use_weighted_sampler: bool = False # TODO: move it to the base config + weighted_sampler_attrs: dict = field(default_factory=lambda: {}) + weighted_sampler_multipliers: dict = field(default_factory=lambda: {}) + + # overrides + r: int = 1 # DO NOT CHANGE + add_blank: bool = True + + # multi-speaker settings + # use speaker embedding layer + num_speakers: int = 0 + speakers_file: Optional[str] = None + speaker_embedding_channels: int = 256 + + # use d-vectors + use_d_vector_file: bool = False + d_vector_file: Optional[list[str]] = None + d_vector_dim: Optional[int] = None + + def __post_init__(self) -> None: + for key, val in self.model_args.items(): + if hasattr(self, key): + self[key] = val diff --git a/TTS/vc/configs/shared_configs.py b/TTS/vc/configs/shared_configs.py index 74164a7444..b2fe63d29d 100644 --- a/TTS/vc/configs/shared_configs.py +++ b/TTS/vc/configs/shared_configs.py @@ -1,7 +1,5 @@ -from dataclasses import asdict, dataclass, field -from typing import Dict, List - -from coqpit import Coqpit, check_argument +from dataclasses import dataclass, field +from typing import List from TTS.config import BaseAudioConfig, BaseDatasetConfig, BaseTrainingConfig diff --git a/TTS/vc/modules/__init__.py b/TTS/vc/layers/__init__.py similarity index 100% rename from TTS/vc/modules/__init__.py rename to TTS/vc/layers/__init__.py diff --git a/TTS/vc/modules/freevc/__init__.py b/TTS/vc/layers/freevc/__init__.py similarity index 100% rename from TTS/vc/modules/freevc/__init__.py rename to TTS/vc/layers/freevc/__init__.py diff --git a/TTS/vc/modules/freevc/commons.py b/TTS/vc/layers/freevc/commons.py similarity index 73% rename from TTS/vc/modules/freevc/commons.py rename to TTS/vc/layers/freevc/commons.py index e799cc2a5b..49889e4816 100644 --- a/TTS/vc/modules/freevc/commons.py +++ b/TTS/vc/layers/freevc/commons.py @@ -1,27 +1,17 @@ import math -import numpy as np import torch -from torch import nn from torch.nn import functional as F +from TTS.tts.utils.helpers import convert_pad_shape -def init_weights(m, mean=0.0, std=0.01): + +def init_weights(m: torch.nn.Module, mean: float = 0.0, std: float = 0.01) -> None: classname = m.__class__.__name__ if classname.find("Conv") != -1: m.weight.data.normal_(mean, std) -def get_padding(kernel_size, dilation=1): - return int((kernel_size * dilation - dilation) / 2) - - -def convert_pad_shape(pad_shape): - l = pad_shape[::-1] - pad_shape = [item for sublist in l for item in sublist] - return pad_shape - - def intersperse(lst, item): result = [item] * (len(lst) * 2 + 1) result[1::2] = lst @@ -106,46 +96,11 @@ def subsequent_mask(length): return mask -@torch.jit.script -def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels): - n_channels_int = n_channels[0] - in_act = input_a + input_b - t_act = torch.tanh(in_act[:, :n_channels_int, :]) - s_act = torch.sigmoid(in_act[:, n_channels_int:, :]) - acts = t_act * s_act - return acts - - def shift_1d(x): x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1] return x -def sequence_mask(length, max_length=None): - if max_length is None: - max_length = length.max() - x = torch.arange(max_length, dtype=length.dtype, device=length.device) - return x.unsqueeze(0) < length.unsqueeze(1) - - -def generate_path(duration, mask): - """ - duration: [b, 1, t_x] - mask: [b, 1, t_y, t_x] - """ - device = duration.device - - b, _, t_y, t_x = mask.shape - cum_duration = torch.cumsum(duration, -1) - - cum_duration_flat = cum_duration.view(b * t_x) - path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype) - path = path.view(b, t_x, t_y) - path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1] - path = path.unsqueeze(1).transpose(2, 3) * mask - return path - - def clip_grad_value_(parameters, clip_value, norm_type=2): if isinstance(parameters, torch.Tensor): parameters = [parameters] diff --git a/TTS/vc/layers/freevc/mel_processing.py b/TTS/vc/layers/freevc/mel_processing.py new file mode 100644 index 0000000000..017d900284 --- /dev/null +++ b/TTS/vc/layers/freevc/mel_processing.py @@ -0,0 +1,58 @@ +import logging + +import torch +import torch.utils.data +from librosa.filters import mel as librosa_mel_fn + +from TTS.utils.audio.torch_transforms import amp_to_db + +logger = logging.getLogger(__name__) + +MAX_WAV_VALUE = 32768.0 + +mel_basis = {} +hann_window = {} + + +def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False): + if torch.min(y) < -1.0: + logger.info("Min value is: %.3f", torch.min(y)) + if torch.max(y) > 1.0: + logger.info("Max value is: %.3f", torch.max(y)) + + global mel_basis, hann_window + dtype_device = str(y.dtype) + "_" + str(y.device) + fmax_dtype_device = str(fmax) + "_" + dtype_device + wnsize_dtype_device = str(win_size) + "_" + dtype_device + if fmax_dtype_device not in mel_basis: + mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax) + mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device) + if wnsize_dtype_device not in hann_window: + hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device) + + y = torch.nn.functional.pad( + y.unsqueeze(1), (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)), mode="reflect" + ) + y = y.squeeze(1) + + spec = torch.view_as_real( + torch.stft( + y, + n_fft, + hop_length=hop_size, + win_length=win_size, + window=hann_window[wnsize_dtype_device], + center=center, + pad_mode="reflect", + normalized=False, + onesided=True, + return_complex=True, + ) + ) + + spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) + + spec = torch.matmul(mel_basis[fmax_dtype_device], spec) + spec = amp_to_db(spec) + + return spec diff --git a/TTS/vc/modules/freevc/modules.py b/TTS/vc/layers/freevc/modules.py similarity index 79% rename from TTS/vc/modules/freevc/modules.py rename to TTS/vc/layers/freevc/modules.py index 9bb5499003..c34f22d701 100644 --- a/TTS/vc/modules/freevc/modules.py +++ b/TTS/vc/layers/freevc/modules.py @@ -5,27 +5,14 @@ from torch.nn.utils.parametrizations import weight_norm from torch.nn.utils.parametrize import remove_parametrizations -import TTS.vc.modules.freevc.commons as commons -from TTS.vc.modules.freevc.commons import get_padding, init_weights +from TTS.tts.layers.generic.normalization import LayerNorm2 +from TTS.tts.layers.generic.wavenet import fused_add_tanh_sigmoid_multiply +from TTS.vc.layers.freevc.commons import init_weights +from TTS.vocoder.models.hifigan_generator import get_padding LRELU_SLOPE = 0.1 -class LayerNorm(nn.Module): - def __init__(self, channels, eps=1e-5): - super().__init__() - self.channels = channels - self.eps = eps - - self.gamma = nn.Parameter(torch.ones(channels)) - self.beta = nn.Parameter(torch.zeros(channels)) - - def forward(self, x): - x = x.transpose(1, -1) - x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps) - return x.transpose(1, -1) - - class ConvReluNorm(nn.Module): def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout): super().__init__() @@ -40,11 +27,11 @@ def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_la self.conv_layers = nn.ModuleList() self.norm_layers = nn.ModuleList() self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size // 2)) - self.norm_layers.append(LayerNorm(hidden_channels)) + self.norm_layers.append(LayerNorm2(hidden_channels)) self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout)) for _ in range(n_layers - 1): self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size // 2)) - self.norm_layers.append(LayerNorm(hidden_channels)) + self.norm_layers.append(LayerNorm2(hidden_channels)) self.proj = nn.Conv1d(hidden_channels, out_channels, 1) self.proj.weight.data.zero_() self.proj.bias.data.zero_() @@ -59,48 +46,6 @@ def forward(self, x, x_mask): return x * x_mask -class DDSConv(nn.Module): - """ - Dialted and Depth-Separable Convolution - """ - - def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0): - super().__init__() - self.channels = channels - self.kernel_size = kernel_size - self.n_layers = n_layers - self.p_dropout = p_dropout - - self.drop = nn.Dropout(p_dropout) - self.convs_sep = nn.ModuleList() - self.convs_1x1 = nn.ModuleList() - self.norms_1 = nn.ModuleList() - self.norms_2 = nn.ModuleList() - for i in range(n_layers): - dilation = kernel_size**i - padding = (kernel_size * dilation - dilation) // 2 - self.convs_sep.append( - nn.Conv1d(channels, channels, kernel_size, groups=channels, dilation=dilation, padding=padding) - ) - self.convs_1x1.append(nn.Conv1d(channels, channels, 1)) - self.norms_1.append(LayerNorm(channels)) - self.norms_2.append(LayerNorm(channels)) - - def forward(self, x, x_mask, g=None): - if g is not None: - x = x + g - for i in range(self.n_layers): - y = self.convs_sep[i](x * x_mask) - y = self.norms_1[i](y) - y = F.gelu(y) - y = self.convs_1x1[i](y) - y = self.norms_2[i](y) - y = F.gelu(y) - y = self.drop(y) - x = x + y - return x * x_mask - - class WN(torch.nn.Module): def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0): super(WN, self).__init__() @@ -154,7 +99,7 @@ def forward(self, x, x_mask, g=None, **kwargs): else: g_l = torch.zeros_like(x_in) - acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor) + acts = fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor) acts = self.drop(acts) res_skip_acts = self.res_skip_layers[i](acts) @@ -317,24 +262,6 @@ def forward(self, x, *args, reverse=False, **kwargs): return x -class ElementwiseAffine(nn.Module): - def __init__(self, channels): - super().__init__() - self.channels = channels - self.m = nn.Parameter(torch.zeros(channels, 1)) - self.logs = nn.Parameter(torch.zeros(channels, 1)) - - def forward(self, x, x_mask, reverse=False, **kwargs): - if not reverse: - y = self.m + torch.exp(self.logs) * x - y = y * x_mask - logdet = torch.sum(self.logs * x_mask, [1, 2]) - return y, logdet - else: - x = (x - self.m) * torch.exp(-self.logs) * x_mask - return x - - class ResidualCouplingLayer(nn.Module): def __init__( self, diff --git a/TTS/vc/modules/freevc/speaker_encoder/__init__.py b/TTS/vc/layers/freevc/speaker_encoder/__init__.py similarity index 100% rename from TTS/vc/modules/freevc/speaker_encoder/__init__.py rename to TTS/vc/layers/freevc/speaker_encoder/__init__.py diff --git a/TTS/vc/modules/freevc/speaker_encoder/audio.py b/TTS/vc/layers/freevc/speaker_encoder/audio.py similarity index 93% rename from TTS/vc/modules/freevc/speaker_encoder/audio.py rename to TTS/vc/layers/freevc/speaker_encoder/audio.py index 52f6fd0893..5fa317ce45 100644 --- a/TTS/vc/modules/freevc/speaker_encoder/audio.py +++ b/TTS/vc/layers/freevc/speaker_encoder/audio.py @@ -1,13 +1,17 @@ -import struct from pathlib import Path from typing import Optional, Union # import webrtcvad import librosa import numpy as np -from scipy.ndimage.morphology import binary_dilation -from TTS.vc.modules.freevc.speaker_encoder.hparams import * +from TTS.vc.layers.freevc.speaker_encoder.hparams import ( + audio_norm_target_dBFS, + mel_n_channels, + mel_window_length, + mel_window_step, + sampling_rate, +) int16_max = (2**15) - 1 diff --git a/TTS/vc/modules/freevc/speaker_encoder/hparams.py b/TTS/vc/layers/freevc/speaker_encoder/hparams.py similarity index 100% rename from TTS/vc/modules/freevc/speaker_encoder/hparams.py rename to TTS/vc/layers/freevc/speaker_encoder/hparams.py diff --git a/TTS/vc/modules/freevc/speaker_encoder/speaker_encoder.py b/TTS/vc/layers/freevc/speaker_encoder/speaker_encoder.py similarity index 94% rename from TTS/vc/modules/freevc/speaker_encoder/speaker_encoder.py rename to TTS/vc/layers/freevc/speaker_encoder/speaker_encoder.py index 2e21a14fd8..a6d5bcf942 100644 --- a/TTS/vc/modules/freevc/speaker_encoder/speaker_encoder.py +++ b/TTS/vc/layers/freevc/speaker_encoder/speaker_encoder.py @@ -1,18 +1,28 @@ -from pathlib import Path +import logging from time import perf_counter as timer from typing import List, Union import numpy as np import torch from torch import nn +from trainer.io import load_fsspec -from TTS.utils.io import load_fsspec -from TTS.vc.modules.freevc.speaker_encoder import audio -from TTS.vc.modules.freevc.speaker_encoder.hparams import * +from TTS.vc.layers.freevc.speaker_encoder import audio +from TTS.vc.layers.freevc.speaker_encoder.hparams import ( + mel_n_channels, + mel_window_step, + model_embedding_size, + model_hidden_size, + model_num_layers, + partials_n_frames, + sampling_rate, +) + +logger = logging.getLogger(__name__) class SpeakerEncoder(nn.Module): - def __init__(self, weights_fpath, device: Union[str, torch.device] = None, verbose=True): + def __init__(self, weights_fpath, device: Union[str, torch.device] = None): """ :param device: either a torch device or the name of a torch device (e.g. "cpu", "cuda"). If None, defaults to cuda if it is available on your machine, otherwise the model will @@ -43,9 +53,7 @@ def __init__(self, weights_fpath, device: Union[str, torch.device] = None, verbo self.load_state_dict(checkpoint["model_state"], strict=False) self.to(device) - - if verbose: - print("Loaded the voice encoder model on %s in %.2f seconds." % (device.type, timer() - start)) + logger.info("Loaded the voice encoder model on %s in %.2f seconds.", device.type, timer() - start) def forward(self, mels: torch.FloatTensor): """ diff --git a/TTS/vc/modules/freevc/wavlm/__init__.py b/TTS/vc/layers/freevc/wavlm/__init__.py similarity index 71% rename from TTS/vc/modules/freevc/wavlm/__init__.py rename to TTS/vc/layers/freevc/wavlm/__init__.py index 6edada407b..62f7e74aaf 100644 --- a/TTS/vc/modules/freevc/wavlm/__init__.py +++ b/TTS/vc/layers/freevc/wavlm/__init__.py @@ -1,10 +1,14 @@ +import logging import os import urllib.request import torch +from trainer.io import get_user_data_dir -from TTS.utils.generic_utils import get_user_data_dir -from TTS.vc.modules.freevc.wavlm.wavlm import WavLM, WavLMConfig +from TTS.utils.generic_utils import is_pytorch_at_least_2_4 +from TTS.vc.layers.freevc.wavlm.wavlm import WavLM, WavLMConfig + +logger = logging.getLogger(__name__) model_uri = "https://github.com/coqui-ai/TTS/releases/download/v0.13.0_models/WavLM-Large.pt" @@ -20,10 +24,10 @@ def get_wavlm(device="cpu"): output_path = os.path.join(output_path, "WavLM-Large.pt") if not os.path.exists(output_path): - print(f" > Downloading WavLM model to {output_path} ...") + logger.info("Downloading WavLM model to %s ...", output_path) urllib.request.urlretrieve(model_uri, output_path) - checkpoint = torch.load(output_path, map_location=torch.device(device)) + checkpoint = torch.load(output_path, map_location=torch.device(device), weights_only=is_pytorch_at_least_2_4()) cfg = WavLMConfig(checkpoint["cfg"]) wavlm = WavLM(cfg).to(device) wavlm.load_state_dict(checkpoint["model"]) diff --git a/TTS/vc/modules/freevc/wavlm/config.json b/TTS/vc/layers/freevc/wavlm/config.json similarity index 99% rename from TTS/vc/modules/freevc/wavlm/config.json rename to TTS/vc/layers/freevc/wavlm/config.json index c6f851b93d..c2e414cf0b 100644 --- a/TTS/vc/modules/freevc/wavlm/config.json +++ b/TTS/vc/layers/freevc/wavlm/config.json @@ -96,4 +96,4 @@ "transformers_version": "4.15.0.dev0", "use_weighted_layer_sum": false, "vocab_size": 32 - } \ No newline at end of file + } diff --git a/TTS/vc/modules/freevc/wavlm/modules.py b/TTS/vc/layers/freevc/wavlm/modules.py similarity index 100% rename from TTS/vc/modules/freevc/wavlm/modules.py rename to TTS/vc/layers/freevc/wavlm/modules.py diff --git a/TTS/vc/modules/freevc/wavlm/wavlm.py b/TTS/vc/layers/freevc/wavlm/wavlm.py similarity index 97% rename from TTS/vc/modules/freevc/wavlm/wavlm.py rename to TTS/vc/layers/freevc/wavlm/wavlm.py index fc93bd4f50..775f3e5979 100644 --- a/TTS/vc/modules/freevc/wavlm/wavlm.py +++ b/TTS/vc/layers/freevc/wavlm/wavlm.py @@ -17,7 +17,7 @@ import torch.nn.functional as F from torch.nn import LayerNorm -from TTS.vc.modules.freevc.wavlm.modules import ( +from TTS.vc.layers.freevc.wavlm.modules import ( Fp32GroupNorm, Fp32LayerNorm, GLU_Linear, @@ -155,7 +155,9 @@ def arrange(s, e, length, keep_length): class WavLMConfig: def __init__(self, cfg=None): - self.extractor_mode: str = "default" # mode for feature extractor. default has a single group norm with d groups in the first conv block, whereas layer_norm has layer norms in every block (meant to use with normalize=True) + self.extractor_mode: str = ( + "default" # mode for feature extractor. default has a single group norm with d groups in the first conv block, whereas layer_norm has layer norms in every block (meant to use with normalize=True) + ) self.encoder_layers: int = 12 # num encoder layers in the transformer self.encoder_embed_dim: int = 768 # encoder embedding dimension @@ -164,7 +166,9 @@ def __init__(self, cfg=None): self.activation_fn: str = "gelu" # activation function to use self.layer_norm_first: bool = False # apply layernorm first in the transformer - self.conv_feature_layers: str = "[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2" # string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...] + self.conv_feature_layers: str = ( + "[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2" # string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...] + ) self.conv_bias: bool = False # include bias in conv encoder self.feature_grad_mult: float = 1.0 # multiply feature extractor var grads by this @@ -387,7 +391,7 @@ def make_conv(): nn.init.kaiming_normal_(conv.weight) return conv - assert (is_layer_norm and is_group_norm) == False, "layer norm and group norm are exclusive" + assert (is_layer_norm and is_group_norm) is False, "layer norm and group norm are exclusive" if is_layer_norm: return nn.Sequential( diff --git a/TTS/vc/models/__init__.py b/TTS/vc/models/__init__.py index 5a09b4e53e..a9807d7006 100644 --- a/TTS/vc/models/__init__.py +++ b/TTS/vc/models/__init__.py @@ -1,15 +1,13 @@ import importlib +import logging import re from typing import Dict, List, Union - -def to_camel(text): - text = text.capitalize() - return re.sub(r"(?!^)_([a-zA-Z])", lambda m: m.group(1).upper(), text) +logger = logging.getLogger(__name__) def setup_model(config: "Coqpit", samples: Union[List[List], List[Dict]] = None) -> "BaseVC": - print(" > Using model: {}".format(config.model)) + logger.info("Using model: %s", config.model) # fetch the right model implementation. if "model" in config and config["model"].lower() == "freevc": MyModel = importlib.import_module("TTS.vc.models.freevc").FreeVC diff --git a/TTS/vc/models/base_vc.py b/TTS/vc/models/base_vc.py index 19f2761bbc..22ffd0095c 100644 --- a/TTS/vc/models/base_vc.py +++ b/TTS/vc/models/base_vc.py @@ -1,6 +1,7 @@ +import logging import os import random -from typing import Dict, List, Tuple, Union +from typing import Any, Optional, Union import torch import torch.distributed as dist @@ -9,6 +10,7 @@ from torch.utils.data import DataLoader from torch.utils.data.sampler import WeightedRandomSampler from trainer.torch import DistributedSampler, DistributedSamplerWrapper +from trainer.trainer import Trainer from TTS.model import BaseTrainerModel from TTS.tts.datasets.dataset import TTSDataset @@ -17,9 +19,12 @@ from TTS.tts.utils.speakers import SpeakerManager, get_speaker_balancer_weights from TTS.tts.utils.synthesis import synthesis from TTS.tts.utils.visual import plot_alignment, plot_spectrogram +from TTS.utils.audio.processor import AudioProcessor # pylint: skip-file +logger = logging.getLogger(__name__) + class BaseVC(BaseTrainerModel): """Base `vc` class. Every new `vc` model must inherit this. @@ -32,10 +37,10 @@ class BaseVC(BaseTrainerModel): def __init__( self, config: Coqpit, - ap: "AudioProcessor", - speaker_manager: SpeakerManager = None, - language_manager: LanguageManager = None, - ): + ap: AudioProcessor, + speaker_manager: Optional[SpeakerManager] = None, + language_manager: Optional[LanguageManager] = None, + ) -> None: super().__init__() self.config = config self.ap = ap @@ -43,7 +48,7 @@ def __init__( self.language_manager = language_manager self._set_model_args(config) - def _set_model_args(self, config: Coqpit): + def _set_model_args(self, config: Coqpit) -> None: """Setup model args based on the config type (`ModelConfig` or `ModelArgs`). `ModelArgs` has all the fields reuqired to initialize the model architecture. @@ -64,7 +69,7 @@ def _set_model_args(self, config: Coqpit): else: raise ValueError("config must be either a *Config or *Args") - def init_multispeaker(self, config: Coqpit, data: List = None): + def init_multispeaker(self, config: Coqpit, data: Optional[list[Any]] = None) -> None: """Initialize a speaker embedding layer if needen and define expected embedding channel size for defining `in_channels` size of the connected layers. @@ -93,15 +98,15 @@ def init_multispeaker(self, config: Coqpit, data: List = None): ) # init speaker embedding layer if config.use_speaker_embedding and not config.use_d_vector_file: - print(" > Init speaker_embedding layer.") + logger.info("Init speaker_embedding layer.") self.speaker_embedding = nn.Embedding(self.num_speakers, self.embedded_speaker_dim) self.speaker_embedding.weight.data.normal_(0, 0.3) - def get_aux_input(self, **kwargs) -> Dict: + def get_aux_input(self, **kwargs: Any) -> dict[str, Any]: """Prepare and return `aux_input` used by `forward()`""" return {"speaker_id": None, "style_wav": None, "d_vector": None, "language_id": None} - def get_aux_input_from_test_sentences(self, sentence_info): + def get_aux_input_from_test_sentences(self, sentence_info: Union[str, list[str]]) -> dict[str, Any]: if hasattr(self.config, "model_args"): config = self.config.model_args else: @@ -129,7 +134,7 @@ def get_aux_input_from_test_sentences(self, sentence_info): if speaker_name is None: d_vector = self.speaker_manager.get_random_embedding() else: - d_vector = self.speaker_manager.get_d_vector_by_name(speaker_name) + d_vector = self.speaker_manager.get_mean_embedding(speaker_name) elif config.use_speaker_embedding: if speaker_name is None: speaker_id = self.speaker_manager.get_random_id() @@ -148,16 +153,16 @@ def get_aux_input_from_test_sentences(self, sentence_info): "language_id": language_id, } - def format_batch(self, batch: Dict) -> Dict: + def format_batch(self, batch: dict[str, Any]) -> dict[str, Any]: """Generic batch formatting for `VCDataset`. You must override this if you use a custom dataset. Args: - batch (Dict): [description] + batch (dict): [description] Returns: - Dict: [description] + dict: [description] """ # setup input batch text_input = batch["token_id"] @@ -227,18 +232,18 @@ def format_batch(self, batch: Dict) -> Dict: "audio_unique_names": batch["audio_unique_names"], } - def get_sampler(self, config: Coqpit, dataset: TTSDataset, num_gpus=1): + def get_sampler(self, config: Coqpit, dataset: TTSDataset, num_gpus: int = 1): weights = None data_items = dataset.samples if getattr(config, "use_language_weighted_sampler", False): alpha = getattr(config, "language_weighted_sampler_alpha", 1.0) - print(" > Using Language weighted sampler with alpha:", alpha) + logger.info("Using Language weighted sampler with alpha: %.2f", alpha) weights = get_language_balancer_weights(data_items) * alpha if getattr(config, "use_speaker_weighted_sampler", False): alpha = getattr(config, "speaker_weighted_sampler_alpha", 1.0) - print(" > Using Speaker weighted sampler with alpha:", alpha) + logger.info("Using Speaker weighted sampler with alpha: %.2f", alpha) if weights is not None: weights += get_speaker_balancer_weights(data_items) * alpha else: @@ -246,7 +251,7 @@ def get_sampler(self, config: Coqpit, dataset: TTSDataset, num_gpus=1): if getattr(config, "use_length_weighted_sampler", False): alpha = getattr(config, "length_weighted_sampler_alpha", 1.0) - print(" > Using Length weighted sampler with alpha:", alpha) + logger.info("Using Length weighted sampler with alpha: %.2f", alpha) if weights is not None: weights += get_length_balancer_weights(data_items) * alpha else: @@ -268,12 +273,12 @@ def get_sampler(self, config: Coqpit, dataset: TTSDataset, num_gpus=1): def get_data_loader( self, config: Coqpit, - assets: Dict, + assets: dict, is_eval: bool, - samples: Union[List[Dict], List[List]], + samples: Union[list[dict], list[list]], verbose: bool, num_gpus: int, - rank: int = None, + rank: Optional[int] = None, ) -> "DataLoader": if is_eval and not config.run_eval: loader = None @@ -318,7 +323,6 @@ def get_data_loader( phoneme_cache_path=config.phoneme_cache_path, precompute_num_workers=config.precompute_num_workers, use_noise_augment=False if is_eval else config.use_noise_augment, - verbose=verbose, speaker_id_mapping=speaker_id_mapping, d_vector_mapping=d_vector_mapping if config.use_d_vector_file else None, tokenizer=None, @@ -350,22 +354,24 @@ def get_data_loader( def _get_test_aux_input( self, - ) -> Dict: + ) -> dict[str, Any]: d_vector = None - if self.config.use_d_vector_file: + if self.speaker_manager is not None and self.config.use_d_vector_file: d_vector = [self.speaker_manager.embeddings[name]["embedding"] for name in self.speaker_manager.embeddings] d_vector = (random.sample(sorted(d_vector), 1),) aux_inputs = { - "speaker_id": None - if not self.config.use_speaker_embedding - else random.sample(sorted(self.speaker_manager.name_to_id.values()), 1), + "speaker_id": ( + None + if not self.config.use_speaker_embedding + else random.sample(sorted(self.speaker_manager.name_to_id.values()), 1) + ), "d_vector": d_vector, "style_wav": None, # TODO: handle GST style input } return aux_inputs - def test_run(self, assets: Dict) -> Tuple[Dict, Dict]: + def test_run(self, assets: dict) -> tuple[dict, dict]: """Generic test run for `vc` models used by `Trainer`. You can override this for a different behaviour. @@ -374,9 +380,9 @@ def test_run(self, assets: Dict) -> Tuple[Dict, Dict]: assets (dict): A dict of training assets. For `vc` models, it must include `{'audio_processor': ap}`. Returns: - Tuple[Dict, Dict]: Test figures and audios to be projected to Tensorboard. + tuple[dict, dict]: Test figures and audios to be projected to Tensorboard. """ - print(" | > Synthesizing test sentences.") + logger.info("Synthesizing test sentences.") test_audios = {} test_figures = {} test_sentences = self.config.test_sentences @@ -405,7 +411,7 @@ def test_run(self, assets: Dict) -> Tuple[Dict, Dict]: ) return test_figures, test_audios - def on_init_start(self, trainer): + def on_init_start(self, trainer: Trainer) -> None: """Save the speaker.pth and language_ids.json at the beginning of the training. Also update both paths.""" if self.speaker_manager is not None: output_path = os.path.join(trainer.output_path, "speakers.pth") @@ -415,8 +421,8 @@ def on_init_start(self, trainer): if hasattr(trainer.config, "model_args"): trainer.config.model_args.speakers_file = output_path trainer.config.save_json(os.path.join(trainer.output_path, "config.json")) - print(f" > `speakers.pth` is saved to {output_path}.") - print(" > `speakers_file` is updated in the config.json.") + logger.info("`speakers.pth` is saved to %s", output_path) + logger.info("`speakers_file` is updated in the config.json.") if self.language_manager is not None: output_path = os.path.join(trainer.output_path, "language_ids.json") @@ -425,5 +431,5 @@ def on_init_start(self, trainer): if hasattr(trainer.config, "model_args"): trainer.config.model_args.language_ids_file = output_path trainer.config.save_json(os.path.join(trainer.output_path, "config.json")) - print(f" > `language_ids.json` is saved to {output_path}.") - print(" > `language_ids_file` is updated in the config.json.") + logger.info("`language_ids.json` is saved to %s", output_path) + logger.info("`language_ids_file` is updated in the config.json.") diff --git a/TTS/vc/models/freevc.py b/TTS/vc/models/freevc.py index 8bb9989224..c654219c39 100644 --- a/TTS/vc/models/freevc.py +++ b/TTS/vc/models/freevc.py @@ -1,3 +1,4 @@ +import logging from typing import Dict, List, Optional, Tuple, Union import librosa @@ -5,22 +6,25 @@ import torch from coqpit import Coqpit from torch import nn -from torch.nn import Conv1d, Conv2d, ConvTranspose1d +from torch.nn import Conv1d, ConvTranspose1d from torch.nn import functional as F -from torch.nn.utils import spectral_norm from torch.nn.utils.parametrizations import weight_norm from torch.nn.utils.parametrize import remove_parametrizations +from trainer.io import load_fsspec -import TTS.vc.modules.freevc.commons as commons -import TTS.vc.modules.freevc.modules as modules +import TTS.vc.layers.freevc.modules as modules +from TTS.tts.layers.vits.discriminator import DiscriminatorS +from TTS.tts.utils.helpers import sequence_mask from TTS.tts.utils.speakers import SpeakerManager -from TTS.utils.io import load_fsspec from TTS.vc.configs.freevc_config import FreeVCConfig +from TTS.vc.layers.freevc.commons import init_weights, rand_slice_segments +from TTS.vc.layers.freevc.mel_processing import mel_spectrogram_torch +from TTS.vc.layers.freevc.speaker_encoder.speaker_encoder import SpeakerEncoder as SpeakerEncoderEx +from TTS.vc.layers.freevc.wavlm import get_wavlm from TTS.vc.models.base_vc import BaseVC -from TTS.vc.modules.freevc.commons import get_padding, init_weights -from TTS.vc.modules.freevc.mel_processing import mel_spectrogram_torch -from TTS.vc.modules.freevc.speaker_encoder.speaker_encoder import SpeakerEncoder as SpeakerEncoderEx -from TTS.vc.modules.freevc.wavlm import get_wavlm +from TTS.vocoder.models.hifigan_discriminator import DiscriminatorP + +logger = logging.getLogger(__name__) class ResidualCouplingBlock(nn.Module): @@ -77,7 +81,7 @@ def __init__( self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) def forward(self, x, x_lengths, g=None): - x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) + x_mask = torch.unsqueeze(sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) x = self.pre(x) * x_mask x = self.enc(x, x_mask, g=g) stats = self.proj(x) * x_mask @@ -152,82 +156,13 @@ def forward(self, x, g=None): return x def remove_weight_norm(self): - print("Removing weight norm...") + logger.info("Removing weight norm...") for l in self.ups: remove_parametrizations(l, "weight") for l in self.resblocks: remove_parametrizations(l, "weight") -class DiscriminatorP(torch.nn.Module): - def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False): - super(DiscriminatorP, self).__init__() - self.period = period - self.use_spectral_norm = use_spectral_norm - norm_f = weight_norm if use_spectral_norm == False else spectral_norm - self.convs = nn.ModuleList( - [ - norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), - norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), - norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), - norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), - norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))), - ] - ) - self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0))) - - def forward(self, x): - fmap = [] - - # 1d to 2d - b, c, t = x.shape - if t % self.period != 0: # pad first - n_pad = self.period - (t % self.period) - x = F.pad(x, (0, n_pad), "reflect") - t = t + n_pad - x = x.view(b, c, t // self.period, self.period) - - for l in self.convs: - x = l(x) - x = F.leaky_relu(x, modules.LRELU_SLOPE) - fmap.append(x) - x = self.conv_post(x) - fmap.append(x) - x = torch.flatten(x, 1, -1) - - return x, fmap - - -class DiscriminatorS(torch.nn.Module): - def __init__(self, use_spectral_norm=False): - super(DiscriminatorS, self).__init__() - norm_f = weight_norm if use_spectral_norm == False else spectral_norm - self.convs = nn.ModuleList( - [ - norm_f(Conv1d(1, 16, 15, 1, padding=7)), - norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)), - norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)), - norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)), - norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)), - norm_f(Conv1d(1024, 1024, 5, 1, padding=2)), - ] - ) - self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1)) - - def forward(self, x): - fmap = [] - - for l in self.convs: - x = l(x) - x = F.leaky_relu(x, modules.LRELU_SLOPE) - fmap.append(x) - x = self.conv_post(x) - fmap.append(x) - x = torch.flatten(x, 1, -1) - - return x, fmap - - class MultiPeriodDiscriminator(torch.nn.Module): def __init__(self, use_spectral_norm=False): super(MultiPeriodDiscriminator, self).__init__() @@ -377,9 +312,9 @@ def device(self): def load_pretrained_speaker_encoder(self): """Load pretrained speaker encoder model as mentioned in the paper.""" - print(" > Loading pretrained speaker encoder model ...") + logger.info("Loading pretrained speaker encoder model ...") self.enc_spk_ex = SpeakerEncoderEx( - "https://github.com/coqui-ai/TTS/releases/download/v0.13.0_models/speaker_encoder.pt" + "https://github.com/coqui-ai/TTS/releases/download/v0.13.0_models/speaker_encoder.pt", device=self.device ) def init_multispeaker(self, config: Coqpit): @@ -449,7 +384,7 @@ def forward( z_p = self.flow(z, spec_mask, g=g) # Randomly slice z and compute o using dec - z_slice, ids_slice = commons.rand_slice_segments(z, spec_lengths, self.segment_size) + z_slice, ids_slice = rand_slice_segments(z, spec_lengths, self.segment_size) o = self.dec(z_slice, g=g) return o, ids_slice, spec_mask, (z, z_p, m_p, logs_p, m_q, logs_q) @@ -468,7 +403,7 @@ def inference(self, c, g=None, mel=None, c_lengths=None): Returns: torch.Tensor: Output tensor. """ - if c_lengths == None: + if c_lengths is None: c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device) if not self.use_spk: g = self.enc_spk.embed_utterance(mel) @@ -544,11 +479,10 @@ def voice_conversion(self, src, tgt): audio = audio[0][0].data.cpu().float().numpy() return audio - def eval_step(): - ... + def eval_step(): ... @staticmethod - def init_from_config(config: FreeVCConfig, samples: Union[List[List], List[Dict]] = None, verbose=True): + def init_from_config(config: FreeVCConfig, samples: Union[List[List], List[Dict]] = None): model = FreeVC(config) return model @@ -558,5 +492,4 @@ def load_checkpoint(self, config, checkpoint_path, eval=False, strict=True, cach if eval: self.eval() - def train_step(): - ... + def train_step(): ... diff --git a/TTS/vc/models/openvoice.py b/TTS/vc/models/openvoice.py new file mode 100644 index 0000000000..135b0861b9 --- /dev/null +++ b/TTS/vc/models/openvoice.py @@ -0,0 +1,320 @@ +import json +import logging +import os +from pathlib import Path +from typing import Any, Mapping, Optional, Union + +import librosa +import numpy as np +import numpy.typing as npt +import torch +from coqpit import Coqpit +from torch import nn +from torch.nn import functional as F +from trainer.io import load_fsspec + +from TTS.tts.layers.vits.networks import PosteriorEncoder +from TTS.tts.utils.speakers import SpeakerManager +from TTS.utils.audio.torch_transforms import wav_to_spec +from TTS.vc.configs.openvoice_config import OpenVoiceConfig +from TTS.vc.models.base_vc import BaseVC +from TTS.vc.models.freevc import Generator, ResidualCouplingBlock + +logger = logging.getLogger(__name__) + + +class ReferenceEncoder(nn.Module): + """NN module creating a fixed size prosody embedding from a spectrogram. + + inputs: mel spectrograms [batch_size, num_spec_frames, num_mel] + outputs: [batch_size, embedding_dim] + """ + + def __init__(self, spec_channels: int, embedding_dim: int = 0, layernorm: bool = True) -> None: + super().__init__() + self.spec_channels = spec_channels + ref_enc_filters = [32, 32, 64, 64, 128, 128] + K = len(ref_enc_filters) + filters = [1] + ref_enc_filters + convs = [ + torch.nn.utils.parametrizations.weight_norm( + nn.Conv2d( + in_channels=filters[i], + out_channels=filters[i + 1], + kernel_size=(3, 3), + stride=(2, 2), + padding=(1, 1), + ) + ) + for i in range(K) + ] + self.convs = nn.ModuleList(convs) + + out_channels = self.calculate_channels(spec_channels, 3, 2, 1, K) + self.gru = nn.GRU( + input_size=ref_enc_filters[-1] * out_channels, + hidden_size=256 // 2, + batch_first=True, + ) + self.proj = nn.Linear(128, embedding_dim) + self.layernorm = nn.LayerNorm(self.spec_channels) if layernorm else None + + def forward(self, inputs: torch.Tensor) -> torch.Tensor: + N = inputs.size(0) + + out = inputs.view(N, 1, -1, self.spec_channels) # [N, 1, Ty, n_freqs] + if self.layernorm is not None: + out = self.layernorm(out) + + for conv in self.convs: + out = conv(out) + out = F.relu(out) # [N, 128, Ty//2^K, n_mels//2^K] + + out = out.transpose(1, 2) # [N, Ty//2^K, 128, n_mels//2^K] + T = out.size(1) + N = out.size(0) + out = out.contiguous().view(N, T, -1) # [N, Ty//2^K, 128*n_mels//2^K] + + self.gru.flatten_parameters() + _memory, out = self.gru(out) # out --- [1, N, 128] + + return self.proj(out.squeeze(0)) + + def calculate_channels(self, L: int, kernel_size: int, stride: int, pad: int, n_convs: int) -> int: + for _ in range(n_convs): + L = (L - kernel_size + 2 * pad) // stride + 1 + return L + + +class OpenVoice(BaseVC): + """ + OpenVoice voice conversion model (inference only). + + Source: https://github.com/myshell-ai/OpenVoice + Paper: https://arxiv.org/abs/2312.01479 + + Paper abstract: + We introduce OpenVoice, a versatile voice cloning approach that requires + only a short audio clip from the reference speaker to replicate their voice and + generate speech in multiple languages. OpenVoice represents a significant + advancement in addressing the following open challenges in the field: 1) + Flexible Voice Style Control. OpenVoice enables granular control over voice + styles, including emotion, accent, rhythm, pauses, and intonation, in addition + to replicating the tone color of the reference speaker. The voice styles are not + directly copied from and constrained by the style of the reference speaker. + Previous approaches lacked the ability to flexibly manipulate voice styles after + cloning. 2) Zero-Shot Cross-Lingual Voice Cloning. OpenVoice achieves zero-shot + cross-lingual voice cloning for languages not included in the massive-speaker + training set. Unlike previous approaches, which typically require extensive + massive-speaker multi-lingual (MSML) dataset for all languages, OpenVoice can + clone voices into a new language without any massive-speaker training data for + that language. OpenVoice is also computationally efficient, costing tens of + times less than commercially available APIs that offer even inferior + performance. To foster further research in the field, we have made the source + code and trained model publicly accessible. We also provide qualitative results + in our demo website. Prior to its public release, our internal version of + OpenVoice was used tens of millions of times by users worldwide between May and + October 2023, serving as the backend of MyShell. + """ + + def __init__(self, config: Coqpit, speaker_manager: Optional[SpeakerManager] = None) -> None: + super().__init__(config, None, speaker_manager, None) + + self.init_multispeaker(config) + + self.zero_g = self.args.zero_g + self.inter_channels = self.args.inter_channels + self.hidden_channels = self.args.hidden_channels + self.filter_channels = self.args.filter_channels + self.n_heads = self.args.n_heads + self.n_layers = self.args.n_layers + self.kernel_size = self.args.kernel_size + self.p_dropout = self.args.p_dropout + self.resblock = self.args.resblock + self.resblock_kernel_sizes = self.args.resblock_kernel_sizes + self.resblock_dilation_sizes = self.args.resblock_dilation_sizes + self.upsample_rates = self.args.upsample_rates + self.upsample_initial_channel = self.args.upsample_initial_channel + self.upsample_kernel_sizes = self.args.upsample_kernel_sizes + self.n_layers_q = self.args.n_layers_q + self.use_spectral_norm = self.args.use_spectral_norm + self.gin_channels = self.args.gin_channels + self.tau = self.args.tau + + self.spec_channels = config.audio.fft_size // 2 + 1 + + self.dec = Generator( + self.inter_channels, + self.resblock, + self.resblock_kernel_sizes, + self.resblock_dilation_sizes, + self.upsample_rates, + self.upsample_initial_channel, + self.upsample_kernel_sizes, + gin_channels=self.gin_channels, + ) + self.enc_q = PosteriorEncoder( + self.spec_channels, + self.inter_channels, + self.hidden_channels, + kernel_size=5, + dilation_rate=1, + num_layers=16, + cond_channels=self.gin_channels, + ) + + self.flow = ResidualCouplingBlock( + self.inter_channels, + self.hidden_channels, + kernel_size=5, + dilation_rate=1, + n_layers=4, + gin_channels=self.gin_channels, + ) + + self.ref_enc = ReferenceEncoder(self.spec_channels, self.gin_channels) + + @property + def device(self) -> torch.device: + return next(self.parameters()).device + + @staticmethod + def init_from_config(config: OpenVoiceConfig) -> "OpenVoice": + return OpenVoice(config) + + def init_multispeaker(self, config: Coqpit, data: Optional[list[Any]] = None) -> None: + """Initialize multi-speaker modules of a model. A model can be trained either with a speaker embedding layer + or with external `d_vectors` computed from a speaker encoder model. + + You must provide a `speaker_manager` at initialization to set up the multi-speaker modules. + + Args: + config (Coqpit): Model configuration. + data (list, optional): Dataset items to infer number of speakers. Defaults to None. + """ + self.num_spks = config.num_speakers + if self.speaker_manager: + self.num_spks = self.speaker_manager.num_speakers + + def load_checkpoint( + self, + config: OpenVoiceConfig, + checkpoint_path: Union[str, os.PathLike[Any]], + eval: bool = False, + strict: bool = True, + cache: bool = False, + ) -> None: + """Map from OpenVoice's config structure.""" + config_path = Path(checkpoint_path).parent / "config.json" + with open(config_path, encoding="utf-8") as f: + config_org = json.load(f) + self.config.audio.input_sample_rate = config_org["data"]["sampling_rate"] + self.config.audio.output_sample_rate = config_org["data"]["sampling_rate"] + self.config.audio.fft_size = config_org["data"]["filter_length"] + self.config.audio.hop_length = config_org["data"]["hop_length"] + self.config.audio.win_length = config_org["data"]["win_length"] + state = load_fsspec(str(checkpoint_path), map_location=torch.device("cpu"), cache=cache) + self.load_state_dict(state["model"], strict=strict) + if eval: + self.eval() + + def forward(self) -> None: ... + def train_step(self) -> None: ... + def eval_step(self) -> None: ... + + @staticmethod + def _set_x_lengths(x: torch.Tensor, aux_input: Mapping[str, Optional[torch.Tensor]]) -> torch.Tensor: + if "x_lengths" in aux_input and aux_input["x_lengths"] is not None: + return aux_input["x_lengths"] + return torch.tensor(x.shape[1:2]).to(x.device) + + @torch.no_grad() + def inference( + self, + x: torch.Tensor, + aux_input: Mapping[str, Optional[torch.Tensor]] = {"x_lengths": None, "g_src": None, "g_tgt": None}, + ) -> dict[str, torch.Tensor]: + """ + Inference pass of the model + + Args: + x (torch.Tensor): Input tensor. Shape: (batch_size, c_seq_len). + x_lengths (torch.Tensor): Lengths of the input tensor. Shape: (batch_size,). + g_src (torch.Tensor): Source speaker embedding tensor. Shape: (batch_size, spk_emb_dim). + g_tgt (torch.Tensor): Target speaker embedding tensor. Shape: (batch_size, spk_emb_dim). + + Returns: + o_hat: Output spectrogram tensor. Shape: (batch_size, spec_seq_len, spec_dim). + x_mask: Spectrogram mask. Shape: (batch_size, spec_seq_len). + (z, z_p, z_hat): A tuple of latent variables. + """ + x_lengths = self._set_x_lengths(x, aux_input) + if "g_src" in aux_input and aux_input["g_src"] is not None: + g_src = aux_input["g_src"] + else: + raise ValueError("aux_input must define g_src") + if "g_tgt" in aux_input and aux_input["g_tgt"] is not None: + g_tgt = aux_input["g_tgt"] + else: + raise ValueError("aux_input must define g_tgt") + z, _m_q, _logs_q, y_mask = self.enc_q( + x, x_lengths, g=g_src if not self.zero_g else torch.zeros_like(g_src), tau=self.tau + ) + z_p = self.flow(z, y_mask, g=g_src) + z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True) + o_hat = self.dec(z_hat * y_mask, g=g_tgt if not self.zero_g else torch.zeros_like(g_tgt)) + return { + "model_outputs": o_hat, + "y_mask": y_mask, + "z": z, + "z_p": z_p, + "z_hat": z_hat, + } + + def load_audio(self, wav: Union[str, npt.NDArray[np.float32], torch.Tensor, list[float]]) -> torch.Tensor: + """Read and format the input audio.""" + if isinstance(wav, str): + out = torch.from_numpy(librosa.load(wav, sr=self.config.audio.input_sample_rate)[0]) + elif isinstance(wav, np.ndarray): + out = torch.from_numpy(wav) + elif isinstance(wav, list): + out = torch.from_numpy(np.array(wav)) + else: + out = wav + return out.to(self.device).float() + + def extract_se(self, audio: Union[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]: + audio_ref = self.load_audio(audio) + y = torch.FloatTensor(audio_ref) + y = y.to(self.device) + y = y.unsqueeze(0) + spec = wav_to_spec( + y, + n_fft=self.config.audio.fft_size, + hop_length=self.config.audio.hop_length, + win_length=self.config.audio.win_length, + center=False, + ).to(self.device) + with torch.no_grad(): + g = self.ref_enc(spec.transpose(1, 2)).unsqueeze(-1) + + return g, spec + + @torch.inference_mode() + def voice_conversion(self, src: Union[str, torch.Tensor], tgt: Union[str, torch.Tensor]) -> npt.NDArray[np.float32]: + """ + Voice conversion pass of the model. + + Args: + src (str or torch.Tensor): Source utterance. + tgt (str or torch.Tensor): Target utterance. + + Returns: + Output numpy array. + """ + src_se, src_spec = self.extract_se(src) + tgt_se, _ = self.extract_se(tgt) + + aux_input = {"g_src": src_se, "g_tgt": tgt_se} + audio = self.inference(src_spec, aux_input) + return audio["model_outputs"][0, 0].data.cpu().float().numpy() diff --git a/TTS/vc/modules/freevc/mel_processing.py b/TTS/vc/modules/freevc/mel_processing.py deleted file mode 100644 index 2dcbf21493..0000000000 --- a/TTS/vc/modules/freevc/mel_processing.py +++ /dev/null @@ -1,125 +0,0 @@ -import torch -import torch.utils.data -from librosa.filters import mel as librosa_mel_fn - -MAX_WAV_VALUE = 32768.0 - - -def dynamic_range_compression_torch(x, C=1, clip_val=1e-5): - """ - PARAMS - ------ - C: compression factor - """ - return torch.log(torch.clamp(x, min=clip_val) * C) - - -def dynamic_range_decompression_torch(x, C=1): - """ - PARAMS - ------ - C: compression factor used to compress - """ - return torch.exp(x) / C - - -def spectral_normalize_torch(magnitudes): - output = dynamic_range_compression_torch(magnitudes) - return output - - -def spectral_de_normalize_torch(magnitudes): - output = dynamic_range_decompression_torch(magnitudes) - return output - - -mel_basis = {} -hann_window = {} - - -def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False): - if torch.min(y) < -1.0: - print("min value is ", torch.min(y)) - if torch.max(y) > 1.0: - print("max value is ", torch.max(y)) - - global hann_window - dtype_device = str(y.dtype) + "_" + str(y.device) - wnsize_dtype_device = str(win_size) + "_" + dtype_device - if wnsize_dtype_device not in hann_window: - hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device) - - y = torch.nn.functional.pad( - y.unsqueeze(1), (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)), mode="reflect" - ) - y = y.squeeze(1) - - spec = torch.stft( - y, - n_fft, - hop_length=hop_size, - win_length=win_size, - window=hann_window[wnsize_dtype_device], - center=center, - pad_mode="reflect", - normalized=False, - onesided=True, - return_complex=False, - ) - - spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) - return spec - - -def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax): - global mel_basis - dtype_device = str(spec.dtype) + "_" + str(spec.device) - fmax_dtype_device = str(fmax) + "_" + dtype_device - if fmax_dtype_device not in mel_basis: - mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax) - mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device) - spec = torch.matmul(mel_basis[fmax_dtype_device], spec) - spec = spectral_normalize_torch(spec) - return spec - - -def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False): - if torch.min(y) < -1.0: - print("min value is ", torch.min(y)) - if torch.max(y) > 1.0: - print("max value is ", torch.max(y)) - - global mel_basis, hann_window - dtype_device = str(y.dtype) + "_" + str(y.device) - fmax_dtype_device = str(fmax) + "_" + dtype_device - wnsize_dtype_device = str(win_size) + "_" + dtype_device - if fmax_dtype_device not in mel_basis: - mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax) - mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device) - if wnsize_dtype_device not in hann_window: - hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device) - - y = torch.nn.functional.pad( - y.unsqueeze(1), (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)), mode="reflect" - ) - y = y.squeeze(1) - - spec = torch.stft( - y, - n_fft, - hop_length=hop_size, - win_length=win_size, - window=hann_window[wnsize_dtype_device], - center=center, - pad_mode="reflect", - normalized=False, - onesided=True, - return_complex=False, - ) - - spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) - - spec = torch.matmul(mel_basis[fmax_dtype_device], spec) - spec = spectral_normalize_torch(spec) - - return spec diff --git a/TTS/vocoder/datasets/__init__.py b/TTS/vocoder/datasets/__init__.py index 871eb0d202..04462817a8 100644 --- a/TTS/vocoder/datasets/__init__.py +++ b/TTS/vocoder/datasets/__init__.py @@ -10,7 +10,7 @@ from TTS.vocoder.datasets.wavernn_dataset import WaveRNNDataset -def setup_dataset(config: Coqpit, ap: AudioProcessor, is_eval: bool, data_items: List, verbose: bool) -> Dataset: +def setup_dataset(config: Coqpit, ap: AudioProcessor, is_eval: bool, data_items: List) -> Dataset: if config.model.lower() in "gan": dataset = GANDataset( ap=ap, @@ -24,7 +24,6 @@ def setup_dataset(config: Coqpit, ap: AudioProcessor, is_eval: bool, data_items: return_segments=not is_eval, use_noise_augment=config.use_noise_augment, use_cache=config.use_cache, - verbose=verbose, ) dataset.shuffle_mapping() elif config.model.lower() == "wavegrad": @@ -39,7 +38,6 @@ def setup_dataset(config: Coqpit, ap: AudioProcessor, is_eval: bool, data_items: return_segments=True, use_noise_augment=False, use_cache=config.use_cache, - verbose=verbose, ) elif config.model.lower() == "wavernn": dataset = WaveRNNDataset( @@ -51,7 +49,6 @@ def setup_dataset(config: Coqpit, ap: AudioProcessor, is_eval: bool, data_items: mode=config.model_params.mode, mulaw=config.model_params.mulaw, is_training=not is_eval, - verbose=verbose, ) else: raise ValueError(f" [!] Dataset for model {config.model.lower()} cannot be found.") diff --git a/TTS/vocoder/datasets/gan_dataset.py b/TTS/vocoder/datasets/gan_dataset.py index 50c38c4deb..0806c0d496 100644 --- a/TTS/vocoder/datasets/gan_dataset.py +++ b/TTS/vocoder/datasets/gan_dataset.py @@ -28,7 +28,6 @@ def __init__( return_segments=True, use_noise_augment=False, use_cache=False, - verbose=False, ): super().__init__() self.ap = ap @@ -43,7 +42,6 @@ def __init__( self.return_segments = return_segments self.use_cache = use_cache self.use_noise_augment = use_noise_augment - self.verbose = verbose assert seq_len % hop_len == 0, " [!] seq_len has to be a multiple of hop_len." self.feat_frame_len = seq_len // hop_len + (2 * conv_pad) @@ -109,7 +107,6 @@ def load_item(self, idx): if self.compute_feat: # compute features from wav wavpath = self.item_list[idx] - # print(wavpath) if self.use_cache and self.cache[idx] is not None: audio, mel = self.cache[idx] diff --git a/TTS/vocoder/datasets/wavegrad_dataset.py b/TTS/vocoder/datasets/wavegrad_dataset.py index 305fe430e3..6f34bccb7c 100644 --- a/TTS/vocoder/datasets/wavegrad_dataset.py +++ b/TTS/vocoder/datasets/wavegrad_dataset.py @@ -28,7 +28,6 @@ def __init__( return_segments=True, use_noise_augment=False, use_cache=False, - verbose=False, ): super().__init__() self.ap = ap @@ -41,7 +40,6 @@ def __init__( self.return_segments = return_segments self.use_cache = use_cache self.use_noise_augment = use_noise_augment - self.verbose = verbose if return_segments: assert seq_len % hop_len == 0, " [!] seq_len has to be a multiple of hop_len." diff --git a/TTS/vocoder/datasets/wavernn_dataset.py b/TTS/vocoder/datasets/wavernn_dataset.py index a67c5b31a0..4c4f5c48df 100644 --- a/TTS/vocoder/datasets/wavernn_dataset.py +++ b/TTS/vocoder/datasets/wavernn_dataset.py @@ -1,9 +1,13 @@ +import logging + import numpy as np import torch from torch.utils.data import Dataset from TTS.utils.audio.numpy_transforms import mulaw_encode, quantize +logger = logging.getLogger(__name__) + class WaveRNNDataset(Dataset): """ @@ -11,9 +15,7 @@ class WaveRNNDataset(Dataset): and converts them to acoustic features on the fly. """ - def __init__( - self, ap, items, seq_len, hop_len, pad, mode, mulaw, is_training=True, verbose=False, return_segments=True - ): + def __init__(self, ap, items, seq_len, hop_len, pad, mode, mulaw, is_training=True, return_segments=True): super().__init__() self.ap = ap self.compute_feat = not isinstance(items[0], (tuple, list)) @@ -25,7 +27,6 @@ def __init__( self.mode = mode self.mulaw = mulaw self.is_training = is_training - self.verbose = verbose self.return_segments = return_segments assert self.seq_len % self.hop_len == 0 @@ -60,7 +61,7 @@ def load_item(self, index): else: min_audio_len = audio.shape[0] + (2 * self.pad * self.hop_len) if audio.shape[0] < min_audio_len: - print(" [!] Instance is too short! : {}".format(wavpath)) + logger.warning("Instance is too short: %s", wavpath) audio = np.pad(audio, [0, min_audio_len - audio.shape[0] + self.hop_len]) mel = self.ap.melspectrogram(audio) @@ -80,7 +81,7 @@ def load_item(self, index): mel = np.load(feat_path.replace("/quant/", "/mel/")) if mel.shape[-1] < self.mel_len + 2 * self.pad: - print(" [!] Instance is too short! : {}".format(wavpath)) + logger.warning("Instance is too short: %s", wavpath) self.item_list[index] = self.item_list[index + 1] feat_path = self.item_list[index] mel = np.load(feat_path.replace("/quant/", "/mel/")) diff --git a/TTS/vocoder/layers/losses.py b/TTS/vocoder/layers/losses.py index 74cfc7262b..8d4dd725ef 100644 --- a/TTS/vocoder/layers/losses.py +++ b/TTS/vocoder/layers/losses.py @@ -221,7 +221,7 @@ class GeneratorLoss(nn.Module): changing configurations. Args: - C (AttrDict): model configuration. + C (Coqpit): model configuration. """ def __init__(self, C): @@ -298,7 +298,7 @@ def forward( adv_loss = adv_loss + self.hinge_gan_loss_weight * hinge_fake_loss # Feature Matching Loss - if self.use_feat_match_loss and not feats_fake is None: + if self.use_feat_match_loss and feats_fake is not None: feat_match_loss = self.feat_match_loss(feats_fake, feats_real) return_dict["G_feat_match_loss"] = feat_match_loss adv_loss = adv_loss + self.feat_match_loss_weight * feat_match_loss diff --git a/TTS/vocoder/models/__init__.py b/TTS/vocoder/models/__init__.py index 65901617b6..b6a1850484 100644 --- a/TTS/vocoder/models/__init__.py +++ b/TTS/vocoder/models/__init__.py @@ -1,12 +1,12 @@ import importlib +import logging import re from coqpit import Coqpit +from TTS.utils.generic_utils import to_camel -def to_camel(text): - text = text.capitalize() - return re.sub(r"(?!^)_([a-zA-Z])", lambda m: m.group(1).upper(), text) +logger = logging.getLogger(__name__) def setup_model(config: Coqpit): @@ -27,13 +27,13 @@ def setup_model(config: Coqpit): MyModel = getattr(MyModel, to_camel(config.model)) except ModuleNotFoundError as e: raise ValueError(f"Model {config.model} not exist!") from e - print(" > Vocoder Model: {}".format(config.model)) + logger.info("Vocoder model: %s", config.model) return MyModel.init_from_config(config) def setup_generator(c): """TODO: use config object as arguments""" - print(" > Generator Model: {}".format(c.generator_model)) + logger.info("Generator model: %s", c.generator_model) MyModel = importlib.import_module("TTS.vocoder.models." + c.generator_model.lower()) MyModel = getattr(MyModel, to_camel(c.generator_model)) # this is to preserve the Wavernn class name (instead of Wavernn) @@ -96,7 +96,7 @@ def setup_generator(c): def setup_discriminator(c): """TODO: use config objekt as arguments""" - print(" > Discriminator Model: {}".format(c.discriminator_model)) + logger.info("Discriminator model: %s", c.discriminator_model) if "parallel_wavegan" in c.discriminator_model: MyModel = importlib.import_module("TTS.vocoder.models.parallel_wavegan_discriminator") else: diff --git a/TTS/vocoder/models/gan.py b/TTS/vocoder/models/gan.py index 19c30e983e..8792950a56 100644 --- a/TTS/vocoder/models/gan.py +++ b/TTS/vocoder/models/gan.py @@ -7,10 +7,10 @@ from torch import nn from torch.utils.data import DataLoader from torch.utils.data.distributed import DistributedSampler +from trainer.io import load_fsspec from trainer.trainer_utils import get_optimizer, get_scheduler from TTS.utils.audio import AudioProcessor -from TTS.utils.io import load_fsspec from TTS.vocoder.datasets.gan_dataset import GANDataset from TTS.vocoder.layers.losses import DiscriminatorLoss, GeneratorLoss from TTS.vocoder.models import setup_discriminator, setup_generator @@ -349,7 +349,6 @@ def get_data_loader( # pylint: disable=no-self-use, unused-argument return_segments=not is_eval, use_noise_augment=config.use_noise_augment, use_cache=config.use_cache, - verbose=verbose, ) dataset.shuffle_mapping() sampler = DistributedSampler(dataset, shuffle=True) if num_gpus > 1 else None @@ -369,6 +368,6 @@ def get_criterion(self): return [DiscriminatorLoss(self.config), GeneratorLoss(self.config)] @staticmethod - def init_from_config(config: Coqpit, verbose=True) -> "GAN": - ap = AudioProcessor.init_from_config(config, verbose=verbose) + def init_from_config(config: Coqpit) -> "GAN": + ap = AudioProcessor.init_from_config(config) return GAN(config, ap=ap) diff --git a/TTS/vocoder/models/hifigan_discriminator.py b/TTS/vocoder/models/hifigan_discriminator.py index 7447a5fbc4..1cbc6ab357 100644 --- a/TTS/vocoder/models/hifigan_discriminator.py +++ b/TTS/vocoder/models/hifigan_discriminator.py @@ -3,6 +3,8 @@ from torch import nn from torch.nn import functional as F +from TTS.vocoder.models.hifigan_generator import get_padding + LRELU_SLOPE = 0.1 @@ -29,7 +31,6 @@ class DiscriminatorP(torch.nn.Module): def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False): super().__init__() self.period = period - get_padding = lambda k, d: int((k * d - d) / 2) norm_f = nn.utils.spectral_norm if use_spectral_norm else nn.utils.parametrizations.weight_norm self.convs = nn.ModuleList( [ diff --git a/TTS/vocoder/models/hifigan_generator.py b/TTS/vocoder/models/hifigan_generator.py index 9247532259..8273d02037 100644 --- a/TTS/vocoder/models/hifigan_generator.py +++ b/TTS/vocoder/models/hifigan_generator.py @@ -1,18 +1,21 @@ # adopted from https://github.com/jik876/hifi-gan/blob/master/models.py +import logging + import torch from torch import nn from torch.nn import Conv1d, ConvTranspose1d from torch.nn import functional as F from torch.nn.utils.parametrizations import weight_norm from torch.nn.utils.parametrize import remove_parametrizations +from trainer.io import load_fsspec -from TTS.utils.io import load_fsspec +logger = logging.getLogger(__name__) LRELU_SLOPE = 0.1 -def get_padding(k, d): - return int((k * d - d) / 2) +def get_padding(kernel_size: int, dilation: int = 1) -> int: + return int((kernel_size * dilation - dilation) / 2) class ResBlock1(torch.nn.Module): @@ -175,6 +178,7 @@ def __init__( conv_pre_weight_norm=True, conv_post_weight_norm=True, conv_post_bias=True, + cond_in_each_up_layer=False, ): r"""HiFiGAN Generator with Multi-Receptive Field Fusion (MRF) @@ -199,6 +203,8 @@ def __init__( self.inference_padding = inference_padding self.num_kernels = len(resblock_kernel_sizes) self.num_upsamples = len(upsample_factors) + self.cond_in_each_up_layer = cond_in_each_up_layer + # initial upsampling layers self.conv_pre = weight_norm(Conv1d(in_channels, upsample_initial_channel, 7, 1, padding=3)) resblock = ResBlock1 if resblock_type == "1" else ResBlock2 @@ -233,6 +239,12 @@ def __init__( if not conv_post_weight_norm: remove_parametrizations(self.conv_post, "weight") + if self.cond_in_each_up_layer: + self.conds = nn.ModuleList() + for i in range(len(self.ups)): + ch = upsample_initial_channel // (2 ** (i + 1)) + self.conds.append(nn.Conv1d(cond_channels, ch, 1)) + def forward(self, x, g=None): """ Args: @@ -252,6 +264,10 @@ def forward(self, x, g=None): for i in range(self.num_upsamples): o = F.leaky_relu(o, LRELU_SLOPE) o = self.ups[i](o) + + if self.cond_in_each_up_layer: + o = o + self.conds[i](g) + z_sum = None for j in range(self.num_kernels): if z_sum is None: @@ -282,7 +298,7 @@ def inference(self, c): return self.forward(c) def remove_weight_norm(self): - print("Removing weight norm...") + logger.info("Removing weight norm...") for l in self.ups: remove_parametrizations(l, "weight") for l in self.resblocks: diff --git a/TTS/vocoder/models/melgan_generator.py b/TTS/vocoder/models/melgan_generator.py index bb3fee789c..03c971afa4 100644 --- a/TTS/vocoder/models/melgan_generator.py +++ b/TTS/vocoder/models/melgan_generator.py @@ -1,8 +1,8 @@ import torch from torch import nn from torch.nn.utils.parametrizations import weight_norm +from trainer.io import load_fsspec -from TTS.utils.io import load_fsspec from TTS.vocoder.layers.melgan import ResidualStack diff --git a/TTS/vocoder/models/parallel_wavegan_discriminator.py b/TTS/vocoder/models/parallel_wavegan_discriminator.py index d02af75f05..211d45d91c 100644 --- a/TTS/vocoder/models/parallel_wavegan_discriminator.py +++ b/TTS/vocoder/models/parallel_wavegan_discriminator.py @@ -1,3 +1,4 @@ +import logging import math import torch @@ -6,6 +7,8 @@ from TTS.vocoder.layers.parallel_wavegan import ResidualBlock +logger = logging.getLogger(__name__) + class ParallelWaveganDiscriminator(nn.Module): """PWGAN discriminator as in https://arxiv.org/abs/1910.11480. @@ -76,7 +79,7 @@ def _apply_weight_norm(m): def remove_weight_norm(self): def _remove_weight_norm(m): try: - # print(f"Weight norm is removed from {m}.") + logger.info("Weight norm is removed from %s", m) remove_parametrizations(m, "weight") except ValueError: # this module didn't have weight norm return @@ -179,7 +182,7 @@ def _apply_weight_norm(m): def remove_weight_norm(self): def _remove_weight_norm(m): try: - print(f"Weight norm is removed from {m}.") + logger.info("Weight norm is removed from %s", m) remove_parametrizations(m, "weight") except ValueError: # this module didn't have weight norm return diff --git a/TTS/vocoder/models/parallel_wavegan_generator.py b/TTS/vocoder/models/parallel_wavegan_generator.py index 8338d94653..e60c8781f0 100644 --- a/TTS/vocoder/models/parallel_wavegan_generator.py +++ b/TTS/vocoder/models/parallel_wavegan_generator.py @@ -1,13 +1,23 @@ +import logging import math import numpy as np import torch from torch.nn.utils.parametrize import remove_parametrizations +from trainer.io import load_fsspec -from TTS.utils.io import load_fsspec from TTS.vocoder.layers.parallel_wavegan import ResidualBlock from TTS.vocoder.layers.upsample import ConvUpsample +logger = logging.getLogger(__name__) + + +def _get_receptive_field_size(layers, stacks, kernel_size, dilation=lambda x: 2**x): + assert layers % stacks == 0 + layers_per_cycle = layers // stacks + dilations = [dilation(i % layers_per_cycle) for i in range(layers)] + return (kernel_size - 1) * sum(dilations) + 1 + class ParallelWaveganGenerator(torch.nn.Module): """PWGAN generator as in https://arxiv.org/pdf/1910.11480.pdf. @@ -126,7 +136,7 @@ def inference(self, c): def remove_weight_norm(self): def _remove_weight_norm(m): try: - # print(f"Weight norm is removed from {m}.") + logger.info("Weight norm is removed from %s", m) remove_parametrizations(m, "weight") except ValueError: # this module didn't have weight norm return @@ -137,20 +147,13 @@ def apply_weight_norm(self): def _apply_weight_norm(m): if isinstance(m, (torch.nn.Conv1d, torch.nn.Conv2d)): torch.nn.utils.parametrizations.weight_norm(m) - # print(f"Weight norm is applied to {m}.") + logger.info("Weight norm is applied to %s", m) self.apply(_apply_weight_norm) - @staticmethod - def _get_receptive_field_size(layers, stacks, kernel_size, dilation=lambda x: 2**x): - assert layers % stacks == 0 - layers_per_cycle = layers // stacks - dilations = [dilation(i % layers_per_cycle) for i in range(layers)] - return (kernel_size - 1) * sum(dilations) + 1 - @property def receptive_field_size(self): - return self._get_receptive_field_size(self.layers, self.stacks, self.kernel_size) + return _get_receptive_field_size(self.layers, self.stacks, self.kernel_size) def load_checkpoint( self, config, checkpoint_path, eval=False, cache=False diff --git a/TTS/vocoder/models/univnet_generator.py b/TTS/vocoder/models/univnet_generator.py index 5e66b70df8..5d1f817927 100644 --- a/TTS/vocoder/models/univnet_generator.py +++ b/TTS/vocoder/models/univnet_generator.py @@ -1,3 +1,4 @@ +import logging from typing import List import numpy as np @@ -6,6 +7,9 @@ from torch.nn.utils import parametrize from TTS.vocoder.layers.lvc_block import LVCBlock +from TTS.vocoder.models.parallel_wavegan_generator import _get_receptive_field_size + +logger = logging.getLogger(__name__) LRELU_SLOPE = 0.1 @@ -113,7 +117,7 @@ def remove_weight_norm(self): def _remove_weight_norm(m): try: - # print(f"Weight norm is removed from {m}.") + logger.info("Weight norm is removed from %s", m) parametrize.remove_parametrizations(m, "weight") except ValueError: # this module didn't have weight norm return @@ -126,21 +130,14 @@ def apply_weight_norm(self): def _apply_weight_norm(m): if isinstance(m, (torch.nn.Conv1d, torch.nn.Conv2d)): torch.nn.utils.parametrizations.weight_norm(m) - # print(f"Weight norm is applied to {m}.") + logger.info("Weight norm is applied to %s", m) self.apply(_apply_weight_norm) - @staticmethod - def _get_receptive_field_size(layers, stacks, kernel_size, dilation=lambda x: 2**x): - assert layers % stacks == 0 - layers_per_cycle = layers // stacks - dilations = [dilation(i % layers_per_cycle) for i in range(layers)] - return (kernel_size - 1) * sum(dilations) + 1 - @property def receptive_field_size(self): """Return receptive field size.""" - return self._get_receptive_field_size(self.layers, self.stacks, self.kernel_size) + return _get_receptive_field_size(self.layers, self.stacks, self.kernel_size) @torch.no_grad() def inference(self, c): diff --git a/TTS/vocoder/models/wavegrad.py b/TTS/vocoder/models/wavegrad.py index c1166e0914..c49abd2201 100644 --- a/TTS/vocoder/models/wavegrad.py +++ b/TTS/vocoder/models/wavegrad.py @@ -9,9 +9,9 @@ from torch.nn.utils.parametrize import remove_parametrizations from torch.utils.data import DataLoader from torch.utils.data.distributed import DistributedSampler +from trainer.io import load_fsspec from trainer.trainer_utils import get_optimizer, get_scheduler -from TTS.utils.io import load_fsspec from TTS.vocoder.datasets import WaveGradDataset from TTS.vocoder.layers.wavegrad import Conv1d, DBlock, FiLM, UBlock from TTS.vocoder.models.base_vocoder import BaseVocoder @@ -321,7 +321,6 @@ def get_data_loader(self, config: Coqpit, assets: Dict, is_eval: True, samples: return_segments=True, use_noise_augment=False, use_cache=config.use_cache, - verbose=verbose, ) sampler = DistributedSampler(dataset) if num_gpus > 1 else None loader = DataLoader( diff --git a/TTS/vocoder/models/wavernn.py b/TTS/vocoder/models/wavernn.py index 7f74ba3ebf..1847679890 100644 --- a/TTS/vocoder/models/wavernn.py +++ b/TTS/vocoder/models/wavernn.py @@ -10,13 +10,14 @@ from torch import nn from torch.utils.data import DataLoader from torch.utils.data.distributed import DistributedSampler +from trainer.io import load_fsspec from TTS.tts.utils.visual import plot_spectrogram from TTS.utils.audio import AudioProcessor from TTS.utils.audio.numpy_transforms import mulaw_decode -from TTS.utils.io import load_fsspec from TTS.vocoder.datasets.wavernn_dataset import WaveRNNDataset from TTS.vocoder.layers.losses import WaveRNNLoss +from TTS.vocoder.layers.upsample import Stretch2d from TTS.vocoder.models.base_vocoder import BaseVocoder from TTS.vocoder.utils.distribution import sample_from_discretized_mix_logistic, sample_from_gaussian @@ -66,19 +67,6 @@ def forward(self, x): return x -class Stretch2d(nn.Module): - def __init__(self, x_scale, y_scale): - super().__init__() - self.x_scale = x_scale - self.y_scale = y_scale - - def forward(self, x): - b, c, h, w = x.size() - x = x.unsqueeze(-1).unsqueeze(3) - x = x.repeat(1, 1, 1, self.y_scale, 1, self.x_scale) - return x.view(b, c, h * self.y_scale, w * self.x_scale) - - class UpsampleNetwork(nn.Module): def __init__( self, @@ -91,7 +79,7 @@ def __init__( use_aux_net, ): super().__init__() - self.total_scale = np.cumproduct(upsample_scales)[-1] + self.total_scale = np.cumprod(upsample_scales)[-1] self.indent = pad * self.total_scale self.use_aux_net = use_aux_net if use_aux_net: @@ -239,7 +227,7 @@ class of models has however remained an elusive problem. With a focus on text-to if self.args.use_upsample_net: assert ( - np.cumproduct(self.args.upsample_factors)[-1] == config.audio.hop_length + np.cumprod(self.args.upsample_factors)[-1] == config.audio.hop_length ), " [!] upsample scales needs to be equal to hop_length" self.upsample = UpsampleNetwork( self.args.feat_dims, @@ -623,7 +611,6 @@ def get_data_loader( # pylint: disable=no-self-use mode=config.model_args.mode, mulaw=config.model_args.mulaw, is_training=not is_eval, - verbose=verbose, ) sampler = DistributedSampler(dataset, shuffle=True) if num_gpus > 1 else None loader = DataLoader( diff --git a/TTS/vocoder/utils/generic_utils.py b/TTS/vocoder/utils/generic_utils.py index 63a0af4445..ac797d97f7 100644 --- a/TTS/vocoder/utils/generic_utils.py +++ b/TTS/vocoder/utils/generic_utils.py @@ -1,3 +1,4 @@ +import logging from typing import Dict import numpy as np @@ -7,6 +8,8 @@ from TTS.tts.utils.visual import plot_spectrogram from TTS.utils.audio import AudioProcessor +logger = logging.getLogger(__name__) + def interpolate_vocoder_input(scale_factor, spec): """Interpolate spectrogram by the scale factor. @@ -20,12 +23,12 @@ def interpolate_vocoder_input(scale_factor, spec): Returns: torch.tensor: interpolated spectrogram. """ - print(" > before interpolation :", spec.shape) + logger.info("Before interpolation: %s", spec.shape) spec = torch.tensor(spec).unsqueeze(0).unsqueeze(0) # pylint: disable=not-callable spec = torch.nn.functional.interpolate( spec, scale_factor=scale_factor, recompute_scale_factor=True, mode="bilinear", align_corners=False ).squeeze(0) - print(" > after interpolation :", spec.shape) + logger.info("After interpolation: %s", spec.shape) return spec @@ -40,7 +43,7 @@ def plot_results(y_hat: torch.tensor, y: torch.tensor, ap: AudioProcessor, name_ Returns: Dict: output figures keyed by the name of the figures. - """ """Plot vocoder model results""" + """ if name_prefix is None: name_prefix = "" diff --git a/dockerfiles/Dockerfile.dev b/dockerfiles/Dockerfile.dev index 58baee53e2..b61bc4de94 100644 --- a/dockerfiles/Dockerfile.dev +++ b/dockerfiles/Dockerfile.dev @@ -11,34 +11,13 @@ RUN apt-get install -y --no-install-recommends \ && rm -rf /var/lib/apt/lists/* # Install Major Python Dependencies: +RUN pip3 install -U pip setuptools RUN pip3 install llvmlite --ignore-installed RUN pip3 install torch torchaudio --extra-index-url https://download.pytorch.org/whl/cu118 RUN rm -rf /root/.cache/pip -WORKDIR /root - -# Copy Dependency Lock Files: -COPY \ - Makefile \ - pyproject.toml \ - setup.py \ - requirements.dev.txt \ - requirements.ja.txt \ - requirements.notebooks.txt \ - requirements.txt \ - /root/ - -# Install Project Dependencies -# Separate stage to limit re-downloading: -RUN pip install \ - -r requirements.txt \ - -r requirements.dev.txt \ - -r requirements.ja.txt \ - -r requirements.notebooks.txt - # Copy TTS repository contents: +WORKDIR /root COPY . /root -# Installing the TTS package itself: -RUN make install - +RUN pip3 install -e .[all,dev] diff --git a/docs/README.md b/docs/README.md deleted file mode 100644 index e69de29bb2..0000000000 diff --git a/docs/requirements.txt b/docs/requirements.txt deleted file mode 100644 index efbefec44b..0000000000 --- a/docs/requirements.txt +++ /dev/null @@ -1,6 +0,0 @@ -furo -myst-parser == 2.0.0 -sphinx == 7.2.5 -sphinx_inline_tabs -sphinx_copybutton -linkify-it-py \ No newline at end of file diff --git a/docs/source/conf.py b/docs/source/conf.py index b85324fd40..e878d0e8f9 100644 --- a/docs/source/conf.py +++ b/docs/source/conf.py @@ -10,26 +10,24 @@ # add these directories to sys.path here. If the directory is relative to the # documentation root, use os.path.abspath to make it absolute, like shown here. # +import importlib.metadata import os import sys -sys.path.insert(0, os.path.abspath('../..')) +sys.path.insert(0, os.path.abspath("../..")) # mock deps with system level requirements. autodoc_mock_imports = ["soundfile"] # -- Project information ----------------------------------------------------- -project = 'TTS' +project = "coqui-tts" copyright = "2021 Coqui GmbH, 2020 TTS authors" -author = 'Coqui GmbH' - -with open("../../TTS/VERSION", "r") as ver: - version = ver.read().strip() +author = "Coqui GmbH" # The version info for the project you're documenting, acts as replacement for # |version| and |release|, also used in various other places throughout the # built documents. -release = version +release = importlib.metadata.version(project) # The main toctree document. master_doc = "index" @@ -40,32 +38,37 @@ # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom # ones. extensions = [ - 'sphinx.ext.autodoc', - 'sphinx.ext.autosummary', - 'sphinx.ext.doctest', - 'sphinx.ext.intersphinx', - 'sphinx.ext.todo', - 'sphinx.ext.coverage', - 'sphinx.ext.napoleon', - 'sphinx.ext.viewcode', - 'sphinx.ext.autosectionlabel', - 'myst_parser', + "sphinx.ext.autodoc", + "sphinx.ext.autosummary", + "sphinx.ext.doctest", + "sphinx.ext.intersphinx", + "sphinx.ext.todo", + "sphinx.ext.coverage", + "sphinx.ext.napoleon", + "sphinx.ext.viewcode", + "sphinx.ext.autosectionlabel", + "myst_parser", "sphinx_copybutton", "sphinx_inline_tabs", ] +suppress_warnings = ["autosectionlabel.*"] # Add any paths that contain templates here, relative to this directory. -templates_path = ['_templates'] +templates_path = ["_templates"] # List of patterns, relative to source directory, that match files and # directories to ignore when looking for source files. # This pattern also affects html_static_path and html_extra_path. -exclude_patterns = ['_build', 'Thumbs.db', '.DS_Store', 'TODO/*'] +exclude_patterns = ["_build", "Thumbs.db", ".DS_Store", "TODO/*"] source_suffix = [".rst", ".md"] -myst_enable_extensions = ['linkify',] +myst_enable_extensions = [ + "linkify", +] + +myst_heading_anchors = 4 # 'sphinxcontrib.katex', # 'sphinx.ext.autosectionlabel', @@ -76,17 +79,17 @@ # duplicated section names that are in different documents. autosectionlabel_prefix_document = True -language = 'en' +language = "en" autodoc_inherit_docstrings = False # Disable displaying type annotations, these can be very verbose -autodoc_typehints = 'none' +autodoc_typehints = "none" # Enable overriding of function signatures in the first line of the docstring. autodoc_docstring_signature = True -napoleon_custom_sections = [('Shapes', 'shape')] +napoleon_custom_sections = [("Shapes", "shape")] # -- Options for HTML output ------------------------------------------------- @@ -94,7 +97,7 @@ # The theme to use for HTML and HTML Help pages. See the documentation for # a list of builtin themes. # -html_theme = 'furo' +html_theme = "furo" html_tite = "TTS" html_theme_options = { "light_logo": "logo.png", @@ -103,18 +106,18 @@ } html_sidebars = { - '**': [ - "sidebar/scroll-start.html", - "sidebar/brand.html", - "sidebar/search.html", - "sidebar/navigation.html", - "sidebar/ethical-ads.html", - "sidebar/scroll-end.html", - ] - } + "**": [ + "sidebar/scroll-start.html", + "sidebar/brand.html", + "sidebar/search.html", + "sidebar/navigation.html", + "sidebar/ethical-ads.html", + "sidebar/scroll-end.html", + ] +} # Add any paths that contain custom static files (such as style sheets) here, # relative to this directory. They are copied after the builtin static files, # so a file named "default.css" will overwrite the builtin "default.css". -html_static_path = ['_static'] +html_static_path = ["_static"] diff --git a/docs/source/configuration.md b/docs/source/configuration.md index ada61e16db..220c96c363 100644 --- a/docs/source/configuration.md +++ b/docs/source/configuration.md @@ -1,6 +1,6 @@ # Configuration -We use 👩‍✈ī¸[Coqpit] for configuration management. It provides basic static type checking and serialization capabilities on top of native Python `dataclasses`. Here is how a simple configuration looks like with Coqpit. +We use 👩‍✈ī¸[Coqpit](https://github.com/idiap/coqui-ai-coqpit) for configuration management. It provides basic static type checking and serialization capabilities on top of native Python `dataclasses`. Here is how a simple configuration looks like with Coqpit. ```python from dataclasses import asdict, dataclass, field @@ -36,7 +36,7 @@ class SimpleConfig(Coqpit): check_argument("val_c", c, restricted=True) ``` -In TTS, each model must have a configuration class that exposes all the values necessary for its lifetime. +In Coqui, each model must have a configuration class that exposes all the values necessary for its lifetime. It defines model architecture, hyper-parameters, training, and inference settings. For our models, we merge all the fields in a single configuration class for ease. It may not look like a wise practice but enables easier bookkeeping and reproducible experiments. diff --git a/docs/source/formatting_your_dataset.md b/docs/source/datasets/formatting_your_dataset.md similarity index 95% rename from docs/source/formatting_your_dataset.md rename to docs/source/datasets/formatting_your_dataset.md index 23c497d0bf..e92263339e 100644 --- a/docs/source/formatting_your_dataset.md +++ b/docs/source/datasets/formatting_your_dataset.md @@ -1,7 +1,9 @@ (formatting_your_dataset)= -# Formatting Your Dataset +# Formatting your dataset -For training a TTS model, you need a dataset with speech recordings and transcriptions. The speech must be divided into audio clips and each clip needs transcription. +For training a TTS model, you need a dataset with speech recordings and +transcriptions. The speech must be divided into audio clips and each clip needs +a transcription. If you have a single audio file and you need to split it into clips, there are different open-source tools for you. We recommend Audacity. It is an open-source and free audio editing software. @@ -49,7 +51,7 @@ The format above is taken from widely-used the [LJSpeech](https://keithito.com/L Your dataset should have good coverage of the target language. It should cover the phonemic variety, exceptional sounds and syllables. This is extremely important for especially non-phonemic languages like English. -For more info about dataset qualities and properties check our [post](https://github.com/coqui-ai/TTS/wiki/What-makes-a-good-TTS-dataset). +For more info about dataset qualities and properties check [this page](what_makes_a_good_dataset.md). ## Using Your Dataset in 🐸TTS diff --git a/docs/source/datasets/index.md b/docs/source/datasets/index.md new file mode 100644 index 0000000000..6b040fc416 --- /dev/null +++ b/docs/source/datasets/index.md @@ -0,0 +1,12 @@ +# Datasets + +For training a TTS model, you need a dataset with speech recordings and +transcriptions. See the following pages for more information on: + +```{toctree} +:maxdepth: 1 + +formatting_your_dataset +what_makes_a_good_dataset +tts_datasets +``` diff --git a/docs/source/tts_datasets.md b/docs/source/datasets/tts_datasets.md similarity index 90% rename from docs/source/tts_datasets.md rename to docs/source/datasets/tts_datasets.md index 11da1b7688..df8d2f2ad9 100644 --- a/docs/source/tts_datasets.md +++ b/docs/source/datasets/tts_datasets.md @@ -1,6 +1,6 @@ -# TTS Datasets +# Public TTS datasets -Some of the known public datasets that we successfully applied 🐸TTS: +Some of the known public datasets that were successfully used for 🐸TTS: - [English - LJ Speech](https://keithito.com/LJ-Speech-Dataset/) - [English - Nancy](http://www.cstr.ed.ac.uk/projects/blizzard/2011/lessac_blizzard2011/) diff --git a/docs/source/what_makes_a_good_dataset.md b/docs/source/datasets/what_makes_a_good_dataset.md similarity index 97% rename from docs/source/what_makes_a_good_dataset.md rename to docs/source/datasets/what_makes_a_good_dataset.md index 18c87453f7..44a93a39da 100644 --- a/docs/source/what_makes_a_good_dataset.md +++ b/docs/source/datasets/what_makes_a_good_dataset.md @@ -17,4 +17,4 @@ If you like to use a bespoken dataset, you might like to perform a couple of qua * **CheckSpectrograms** is to measure the noise level of the clips and find good audio processing parameters. The noise level might be observed by checking spectrograms. If spectrograms look cluttered, especially in silent parts, this dataset might not be a good candidate for a TTS project. If your voice clips are too noisy in the background, it makes things harder for your model to learn the alignment, and the final result might be different than the voice you are given. If the spectrograms look good, then the next step is to find a good set of audio processing parameters, defined in ```config.json```. In the notebook, you can compare different sets of parameters and see the resynthesis results in relation to the given ground-truth. Find the best parameters that give the best possible synthesis performance. -Another practical detail is the quantization level of the clips. If your dataset has a very high bit-rate, that might cause slow data-load time and consequently slow training. It is better to reduce the sample-rate of your dataset to around 16000-22050. \ No newline at end of file +Another practical detail is the quantization level of the clips. If your dataset has a very high bit-rate, that might cause slow data-load time and consequently slow training. It is better to reduce the sample-rate of your dataset to around 16000-22050. diff --git a/docs/source/docker_images.md b/docs/source/docker_images.md index d08a55837d..042f9f8e7a 100644 --- a/docs/source/docker_images.md +++ b/docs/source/docker_images.md @@ -1,20 +1,20 @@ (docker_images)= -## Docker images +# Docker images We provide docker images to be able to test TTS without having to setup your own environment. -### Using premade images +## Using premade images You can use premade images built automatically from the latest TTS version. -#### CPU version +### CPU version ```bash docker pull ghcr.io/coqui-ai/tts-cpu ``` -#### GPU version +### GPU version ```bash docker pull ghcr.io/coqui-ai/tts ``` -### Building your own image +## Building your own image ```bash docker build -t tts . ``` @@ -32,7 +32,7 @@ For the GPU version, you need to have the latest NVIDIA drivers installed. With `nvidia-smi` you can check the CUDA version supported, it must be >= 11.8 ```bash -docker run --rm --gpus all -v ~/tts-output:/root/tts-output ghcr.io/coqui-ai/tts --text "Hello." --out_path /root/tts-output/hello.wav --use_cuda true +docker run --rm --gpus all -v ~/tts-output:/root/tts-output ghcr.io/coqui-ai/tts --text "Hello." --out_path /root/tts-output/hello.wav --use_cuda ``` ## Start a server @@ -50,7 +50,7 @@ python3 TTS/server/server.py --model_name tts_models/en/vctk/vits ```bash docker run --rm -it -p 5002:5002 --gpus all --entrypoint /bin/bash ghcr.io/coqui-ai/tts python3 TTS/server/server.py --list_models #To get the list of available models -python3 TTS/server/server.py --model_name tts_models/en/vctk/vits --use_cuda true +python3 TTS/server/server.py --model_name tts_models/en/vctk/vits --use_cuda ``` -Click [there](http://[::1]:5002/) and have fun with the server! \ No newline at end of file +Click [there](http://[::1]:5002/) and have fun with the server! diff --git a/docs/source/implementing_a_new_language_frontend.md b/docs/source/extension/implementing_a_new_language_frontend.md similarity index 88% rename from docs/source/implementing_a_new_language_frontend.md rename to docs/source/extension/implementing_a_new_language_frontend.md index 2041352d64..0b3ef59be0 100644 --- a/docs/source/implementing_a_new_language_frontend.md +++ b/docs/source/extension/implementing_a_new_language_frontend.md @@ -1,6 +1,6 @@ -# Implementing a New Language Frontend +# Implementing new language front ends -- Language frontends are located under `TTS.tts.utils.text` +- Language front ends are located under `TTS.tts.utils.text` - Each special language has a separate folder. - Each folder contains all the utilities for processing the text input. - `TTS.tts.utils.text.phonemizers` contains the main phonemizer for a language. This is the class that uses the utilities diff --git a/docs/source/implementing_a_new_model.md b/docs/source/extension/implementing_a_new_model.md similarity index 97% rename from docs/source/implementing_a_new_model.md rename to docs/source/extension/implementing_a_new_model.md index 1bf7a8822e..2521789771 100644 --- a/docs/source/implementing_a_new_model.md +++ b/docs/source/extension/implementing_a_new_model.md @@ -1,4 +1,4 @@ -# Implementing a Model +# Implementing new models 1. Implement layers. @@ -36,7 +36,8 @@ There is also the `callback` interface by which you can manipulate both the model and the `Trainer` states. Callbacks give you an infinite flexibility to add custom behaviours for your model and training routines. - For more details, see {ref}`BaseTTS ` and :obj:`TTS.utils.callbacks`. + For more details, see [BaseTTS](../main_classes/model_api.md#base-tts-model) + and `TTS.utils.callbacks`. 6. Optionally, define `MyModelArgs`. @@ -62,7 +63,7 @@ We love you more when you document your code. ❤ī¸ -# Template 🐸TTS Model implementation +## Template 🐸TTS Model implementation You can start implementing your model by copying the following base class. diff --git a/docs/source/extension/index.md b/docs/source/extension/index.md new file mode 100644 index 0000000000..39c36b632c --- /dev/null +++ b/docs/source/extension/index.md @@ -0,0 +1,14 @@ +# Adding models or languages + +You can extend Coqui by implementing new model architectures or adding front +ends for new languages. See the pages below for more details. The [project +structure](../project_structure.md) and [contribution +guidelines](../contributing.md) may also be helpful. Please open a pull request +with your changes to share back the improvements with the community. + +```{toctree} +:maxdepth: 1 + +implementing_a_new_model +implementing_a_new_language_frontend +``` diff --git a/docs/source/faq.md b/docs/source/faq.md index fa48c4a9fb..4fbd149f00 100644 --- a/docs/source/faq.md +++ b/docs/source/faq.md @@ -1,28 +1,56 @@ -# Humble FAQ -We tried to collect common issues and questions we receive about 🐸TTS. It is worth checking before going deeper. +# FAQ +We tried to collect common issues and questions we receive about 🐸TTS. It is +worth checking before going deeper. -## Errors with a pre-trained model. How can I resolve this? -- Make sure you use the right commit version of 🐸TTS. Each pre-trained model has its corresponding version that needs to be used. It is defined on the model table. -- If it is still problematic, post your problem on [Discussions](https://github.com/coqui-ai/TTS/discussions). Please give as many details as possible (error message, your TTS version, your TTS model and config.json etc.) -- If you feel like it's a bug to be fixed, then prefer Github issues with the same level of scrutiny. +## Using Coqui -## What are the requirements of a good 🐸TTS dataset? -* {ref}`See this page ` +### Where does Coqui store downloaded models? -## How should I choose the right model? +The path to downloaded models is printed when running `tts --list_models`. +Default locations are: + +- **Linux:** `~/.local/share/tts` +- **Mac:** `~/Library/Application Support/tts` +- **Windows:** `C:\Users\\AppData\Local\tts` + +You can change the prefix of this `tts/` folder by setting the `XDG_DATA_HOME` +or `TTS_HOME` environment variables. + +### Errors with a pre-trained model. How can I resolve this? +- Make sure you use the latest version of 🐸TTS. Each pre-trained model is only + supported from a certain minimum version. +- If it is still problematic, post your problem on + [Discussions](https://github.com/idiap/coqui-ai-TTS/discussions). Please give + as many details as possible (error message, your TTS version, your TTS model + and config.json etc.) +- If you feel like it's a bug to be fixed, then prefer Github issues with the + same level of scrutiny. + +## Training Coqui models + +### What are the requirements of a good 🐸TTS dataset? +- [See this page](datasets/what_makes_a_good_dataset.md) + +### How should I choose the right model? - First, train Tacotron. It is smaller and faster to experiment with. If it performs poorly, try Tacotron2. - Tacotron models produce the most natural voice if your dataset is not too noisy. - If both models do not perform well and especially the attention does not align, then try AlignTTS or GlowTTS. - If you need faster models, consider SpeedySpeech, GlowTTS or AlignTTS. Keep in mind that SpeedySpeech requires a pre-trained Tacotron or Tacotron2 model to compute text-to-speech alignments. -## How can I train my own `tts` model? -0. Check your dataset with notebooks in [dataset_analysis](https://github.com/coqui-ai/TTS/tree/master/notebooks/dataset_analysis) folder. Use [this notebook](https://github.com/coqui-ai/TTS/blob/master/notebooks/dataset_analysis/CheckSpectrograms.ipynb) to find the right audio processing parameters. A better set of parameters results in a better audio synthesis. +### How can I train my own `tts` model? + +```{note} XTTS has separate fine-tuning scripts, see [here](models/xtts.md#training). +``` + +0. Check your dataset with notebooks in [dataset_analysis](https://github.com/idiap/coqui-ai-TTS/tree/main/notebooks/dataset_analysis) folder. Use [this notebook](https://github.com/idiap/coqui-ai-TTS/blob/main/notebooks/dataset_analysis/CheckSpectrograms.ipynb) to find the right audio processing parameters. A better set of parameters results in a better audio synthesis. -1. Write your own dataset `formatter` in `datasets/formatters.py` or format your dataset as one of the supported datasets, like LJSpeech. +1. Write your own dataset `formatter` in `datasets/formatters.py` or [format](datasets/formatting_your_dataset) your dataset as one of the supported datasets, like LJSpeech. A `formatter` parses the metadata file and converts a list of training samples. 2. If you have a dataset with a different alphabet than English, you need to set your own character list in the ```config.json```. - - If you use phonemes for training and your language is supported [here](https://github.com/rhasspy/gruut#supported-languages), you don't need to set your character list. + - If you use phonemes for training and your language is supported by + [Espeak](https://github.com/espeak-ng/espeak-ng/blob/master/docs/languages.md) + or [Gruut](https://github.com/rhasspy/gruut#supported-languages), you don't need to set your character list. - You can use `TTS/bin/find_unique_chars.py` to get characters used in your dataset. 3. Write your own text cleaner in ```utils.text.cleaners```. It is not always necessary, except when you have a different alphabet or language-specific requirements. @@ -61,15 +89,16 @@ We tried to collect common issues and questions we receive about 🐸TTS. It is - SingleGPU training: ```CUDA_VISIBLE_DEVICES="0" python train_tts.py --config_path config.json``` - MultiGPU training: ```python3 -m trainer.distribute --gpus "0,1" --script TTS/bin/train_tts.py --config_path config.json``` -**Note:** You can also train your model using pure 🐍 python. Check ```{eval-rst} :ref: 'tutorial_for_nervous_beginners'```. +**Note:** You can also train your model using pure 🐍 python. Check the +[tutorial](tutorial_for_nervous_beginners.md). -## How can I train in a different language? +### How can I train in a different language? - Check steps 2, 3, 4, 5 above. -## How can I train multi-GPUs? +### How can I train multi-GPUs? - Check step 5 above. -## How can I check model performance? +### How can I check model performance? - You can inspect model training and performance using ```tensorboard```. It will show you loss, attention alignment, model output. Go with the order below to measure the model performance. 1. Check ground truth spectrograms. If they do not look as they are supposed to, then check audio processing parameters in ```config.json```. 2. Check train and eval losses and make sure that they all decrease smoothly in time. @@ -84,7 +113,7 @@ We tried to collect common issues and questions we receive about 🐸TTS. It is - 'bidirectional_decoder' is your ultimate savior, but it trains 2x slower and demands 1.5x more GPU memory. - You can also try the other models like AlignTTS or GlowTTS. -## How do I know when to stop training? +### How do I know when to stop training? There is no single objective metric to decide the end of a training since the voice quality is a subjective matter. In our model trainings, we follow these steps; @@ -97,17 +126,17 @@ In our model trainings, we follow these steps; Keep in mind that the approach above only validates the model robustness. It is hard to estimate the voice quality without asking the actual people. The best approach is to pick a set of promising models and run a Mean-Opinion-Score study asking actual people to score the models. -## My model does not learn. How can I debug? +### My model does not learn. How can I debug? - Go over the steps under "How can I check model performance?" -## Attention does not align. How can I make it work? +### Attention does not align. How can I make it work? - Check the 4th step under "How can I check model performance?" -## How can I test a trained model? -- The best way is to use `tts` or `tts-server` commands. For details check {ref}`here `. +### How can I test a trained model? +- The best way is to use `tts` or `tts-server` commands. For details check [here](inference.md). - If you need to code your own ```TTS.utils.synthesizer.Synthesizer``` class. -## My Tacotron model does not stop - I see "Decoder stopped with 'max_decoder_steps" - Stopnet does not work. +### My Tacotron model does not stop - I see "Decoder stopped with 'max_decoder_steps" - Stopnet does not work. - In general, all of the above relates to the `stopnet`. It is the part of the model telling the `decoder` when to stop. - In general, a poor `stopnet` relates to something else that is broken in your model or dataset. Especially the attention module. - One common reason is the silent parts in the audio clips at the beginning and the ending. Check ```trim_db``` value in the config. You can find a better value for your dataset by using ```CheckSpectrogram``` notebook. If this value is too small, too much of the audio will be trimmed. If too big, then too much silence will remain. Both will curtail the `stopnet` performance. diff --git a/docs/source/index.md b/docs/source/index.md index 79993eec76..3a030b4f81 100644 --- a/docs/source/index.md +++ b/docs/source/index.md @@ -1,62 +1,63 @@ +--- +hide-toc: true +--- ```{include} ../../README.md :relative-images: +:end-before: ``` ----- - -# Documentation Content -```{eval-rst} -.. toctree:: - :maxdepth: 2 - :caption: Get started - - tutorial_for_nervous_beginners - installation - faq - contributing - -.. toctree:: - :maxdepth: 2 - :caption: Using 🐸TTS - - inference - docker_images - implementing_a_new_model - implementing_a_new_language_frontend - training_a_model - finetuning - configuration - formatting_your_dataset - what_makes_a_good_dataset - tts_datasets - marytts - -.. toctree:: - :maxdepth: 2 - :caption: Main Classes - - main_classes/trainer_api - main_classes/audio_processor - main_classes/model_api - main_classes/dataset - main_classes/gan - main_classes/speaker_manager - -.. toctree:: - :maxdepth: 2 - :caption: `tts` Models - - models/glow_tts.md - models/vits.md - models/forward_tts.md - models/tacotron1-2.md - models/overflow.md - models/tortoise.md - models/bark.md - models/xtts.md - -.. toctree:: - :maxdepth: 2 - :caption: `vocoder` Models +```{toctree} +:maxdepth: 1 +:caption: Get started +:hidden: + +tutorial_for_nervous_beginners +installation +docker_images +faq +project_structure +contributing +``` + +```{toctree} +:maxdepth: 1 +:caption: Using Coqui +:hidden: + +inference +training/index +extension/index +datasets/index +``` + + +```{toctree} +:maxdepth: 1 +:caption: Main Classes +:hidden: + +configuration +main_classes/trainer_api +main_classes/audio_processor +main_classes/model_api +main_classes/dataset +main_classes/gan +main_classes/speaker_manager +``` + + +```{toctree} +:maxdepth: 1 +:caption: TTS Models +:hidden: + +models/glow_tts.md +models/vits.md +models/forward_tts.md +models/tacotron1-2.md +models/overflow.md +models/tortoise.md +models/bark.md +models/xtts.md ``` diff --git a/docs/source/inference.md b/docs/source/inference.md index 56bccfb5b2..cb7d01fca3 100644 --- a/docs/source/inference.md +++ b/docs/source/inference.md @@ -1,192 +1,21 @@ (synthesizing_speech)= -# Synthesizing Speech +# Synthesizing speech -First, you need to install TTS. We recommend using PyPi. You need to call the command below: +## Overview -```bash -$ pip install TTS -``` - -After the installation, 2 terminal commands are available. - -1. TTS Command Line Interface (CLI). - `tts` -2. Local Demo Server. - `tts-server` -3. In 🐍Python. - `from TTS.api import TTS` - -## On the Commandline - `tts` -![cli.gif](https://github.com/coqui-ai/TTS/raw/main/images/tts_cli.gif) - -After the installation, 🐸TTS provides a CLI interface for synthesizing speech using pre-trained models. You can either use your own model or the release models under 🐸TTS. - -Listing released 🐸TTS models. - -```bash -tts --list_models -``` - -Run a TTS model, from the release models list, with its default vocoder. (Simply copy and paste the full model names from the list as arguments for the command below.) - -```bash -tts --text "Text for TTS" \ - --model_name "///" \ - --out_path folder/to/save/output.wav -``` - -Run a tts and a vocoder model from the released model list. Note that not every vocoder is compatible with every TTS model. - -```bash -tts --text "Text for TTS" \ - --model_name "tts_models///" \ - --vocoder_name "vocoder_models///" \ - --out_path folder/to/save/output.wav -``` - -Run your own TTS model (Using Griffin-Lim Vocoder) - -```bash -tts --text "Text for TTS" \ - --model_path path/to/model.pth \ - --config_path path/to/config.json \ - --out_path folder/to/save/output.wav -``` - -Run your own TTS and Vocoder models - -```bash -tts --text "Text for TTS" \ - --config_path path/to/config.json \ - --model_path path/to/model.pth \ - --out_path folder/to/save/output.wav \ - --vocoder_path path/to/vocoder.pth \ - --vocoder_config_path path/to/vocoder_config.json -``` - -Run a multi-speaker TTS model from the released models list. - -```bash -tts --model_name "tts_models///" --list_speaker_idxs # list the possible speaker IDs. -tts --text "Text for TTS." --out_path output/path/speech.wav --model_name "tts_models///" --speaker_idx "" -``` - -Run a released voice conversion model - -```bash -tts --model_name "voice_conversion///" - --source_wav "my/source/speaker/audio.wav" - --target_wav "my/target/speaker/audio.wav" - --out_path folder/to/save/output.wav -``` - -**Note:** You can use ```./TTS/bin/synthesize.py``` if you prefer running ```tts``` from the TTS project folder. - -## On the Demo Server - `tts-server` - - -![server.gif](https://github.com/coqui-ai/TTS/raw/main/images/demo_server.gif) - -You can boot up a demo 🐸TTS server to run an inference with your models. Note that the server is not optimized for performance -but gives you an easy way to interact with the models. +Coqui TTS provides three main methods for inference: -The demo server provides pretty much the same interface as the CLI command. +1. 🐍Python API +2. TTS command line interface (CLI) +3. [Local demo server](server.md) -```bash -tts-server -h # see the help -tts-server --list_models # list the available models. +```{include} ../../README.md +:start-after: ``` -Run a TTS model, from the release models list, with its default vocoder. -If the model you choose is a multi-speaker TTS model, you can select different speakers on the Web interface and synthesize -speech. - -```bash -tts-server --model_name "///" -``` - -Run a TTS and a vocoder model from the released model list. Note that not every vocoder is compatible with every TTS model. - -```bash -tts-server --model_name "///" \ - --vocoder_name "///" -``` - -## Python 🐸TTS API - -You can run a multi-speaker and multi-lingual model in Python as - -```python -import torch -from TTS.api import TTS - -# Get device -device = "cuda" if torch.cuda.is_available() else "cpu" - -# List available 🐸TTS models -print(TTS().list_models()) - -# Init TTS -tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to(device) - -# Run TTS -# ❗ Since this model is multi-lingual voice cloning model, we must set the target speaker_wav and language -# Text to speech list of amplitude values as output -wav = tts.tts(text="Hello world!", speaker_wav="my/cloning/audio.wav", language="en") -# Text to speech to a file -tts.tts_to_file(text="Hello world!", speaker_wav="my/cloning/audio.wav", language="en", file_path="output.wav") -``` - -#### Here is an example for a single speaker model. - -```python -# Init TTS with the target model name -tts = TTS(model_name="tts_models/de/thorsten/tacotron2-DDC", progress_bar=False) -# Run TTS -tts.tts_to_file(text="Ich bin eine Testnachricht.", file_path=OUTPUT_PATH) -``` - -#### Example voice cloning with YourTTS in English, French and Portuguese: - -```python -tts = TTS(model_name="tts_models/multilingual/multi-dataset/your_tts", progress_bar=False).to("cuda") -tts.tts_to_file("This is voice cloning.", speaker_wav="my/cloning/audio.wav", language="en", file_path="output.wav") -tts.tts_to_file("C'est le clonage de la voix.", speaker_wav="my/cloning/audio.wav", language="fr", file_path="output.wav") -tts.tts_to_file("Isso Ê clonagem de voz.", speaker_wav="my/cloning/audio.wav", language="pt", file_path="output.wav") -``` - -#### Example voice conversion converting speaker of the `source_wav` to the speaker of the `target_wav` - -```python -tts = TTS(model_name="voice_conversion_models/multilingual/vctk/freevc24", progress_bar=False).to("cuda") -tts.voice_conversion_to_file(source_wav="my/source.wav", target_wav="my/target.wav", file_path="output.wav") -``` - -#### Example voice cloning by a single speaker TTS model combining with the voice conversion model. - -This way, you can clone voices by using any model in 🐸TTS. - -```python -tts = TTS("tts_models/de/thorsten/tacotron2-DDC") -tts.tts_with_vc_to_file( - "Wie sage ich auf Italienisch, dass ich dich liebe?", - speaker_wav="target/speaker.wav", - file_path="ouptut.wav" -) -``` - -#### Example text to speech using **Fairseq models in ~1100 languages** đŸ¤¯. -For these models use the following name format: `tts_models//fairseq/vits`. - -You can find the list of language ISO codes [here](https://dl.fbaipublicfiles.com/mms/tts/all-tts-languages.html) and learn about the Fairseq models [here](https://github.com/facebookresearch/fairseq/tree/main/examples/mms). - -```python -from TTS.api import TTS -api = TTS(model_name="tts_models/eng/fairseq/vits").to("cuda") -api.tts_to_file("This is a test.", file_path="output.wav") -# TTS with on the fly voice conversion -api = TTS("tts_models/deu/fairseq/vits") -api.tts_with_vc_to_file( - "Wie sage ich auf Italienisch, dass ich dich liebe?", - speaker_wav="target/speaker.wav", - file_path="ouptut.wav" -) +```{toctree} +:hidden: +server +marytts ``` diff --git a/docs/source/installation.md b/docs/source/installation.md index c4d05361f4..1315395a59 100644 --- a/docs/source/installation.md +++ b/docs/source/installation.md @@ -1,33 +1,6 @@ # Installation -🐸TTS supports python >=3.7 <3.11.0 and tested on Ubuntu 18.10, 19.10, 20.10. - -## Using `pip` - -`pip` is recommended if you want to use 🐸TTS only for inference. - -You can install from PyPI as follows: - -```bash -pip install TTS # from PyPI -``` - -Or install from Github: - -```bash -pip install git+https://github.com/coqui-ai/TTS # from Github -``` - -## Installing From Source - -This is recommended for development and more control over 🐸TTS. - -```bash -git clone https://github.com/coqui-ai/TTS/ -cd TTS -make system-deps # only on Linux systems. -make install +```{include} ../../README.md +:start-after: +:end-before: ``` - -## On Windows -If you are on Windows, 👑@GuyPaddock wrote installation instructions [here](https://stackoverflow.com/questions/66726331/ \ No newline at end of file diff --git a/docs/source/main_classes/audio_processor.md b/docs/source/main_classes/audio_processor.md index 600b0db582..98e94a8789 100644 --- a/docs/source/main_classes/audio_processor.md +++ b/docs/source/main_classes/audio_processor.md @@ -22,4 +22,4 @@ also must inherit or initiate `BaseAudioConfig`. ```{eval-rst} .. autoclass:: TTS.config.shared_configs.BaseAudioConfig :members: -``` \ No newline at end of file +``` diff --git a/docs/source/main_classes/dataset.md b/docs/source/main_classes/dataset.md index 92d381aca5..1566488194 100644 --- a/docs/source/main_classes/dataset.md +++ b/docs/source/main_classes/dataset.md @@ -22,4 +22,4 @@ ```{eval-rst} .. autoclass:: TTS.vocoder.datasets.wavernn_dataset.WaveRNNDataset :members: -``` \ No newline at end of file +``` diff --git a/docs/source/main_classes/gan.md b/docs/source/main_classes/gan.md index 4524b4b5c5..e143f6431e 100644 --- a/docs/source/main_classes/gan.md +++ b/docs/source/main_classes/gan.md @@ -9,4 +9,4 @@ to do its ✨ī¸. ```{eval-rst} .. autoclass:: TTS.vocoder.models.gan.GAN :members: -``` \ No newline at end of file +``` diff --git a/docs/source/main_classes/model_api.md b/docs/source/main_classes/model_api.md index 0e6f2d9427..bb7e9d1a1d 100644 --- a/docs/source/main_classes/model_api.md +++ b/docs/source/main_classes/model_api.md @@ -1,24 +1,24 @@ # Model API Model API provides you a set of functions that easily make your model compatible with the `Trainer`, -`Synthesizer` and `ModelZoo`. +`Synthesizer` and the Coqui Python API. -## Base TTS Model +## Base Trainer Model ```{eval-rst} .. autoclass:: TTS.model.BaseTrainerModel :members: ``` -## Base tts Model +## Base TTS Model ```{eval-rst} .. autoclass:: TTS.tts.models.base_tts.BaseTTS :members: ``` -## Base vocoder Model +## Base Vocoder Model ```{eval-rst} .. autoclass:: TTS.vocoder.models.base_vocoder.BaseVocoder :members: -``` \ No newline at end of file +``` diff --git a/docs/source/main_classes/speaker_manager.md b/docs/source/main_classes/speaker_manager.md index ba4b55dc78..fe98823956 100644 --- a/docs/source/main_classes/speaker_manager.md +++ b/docs/source/main_classes/speaker_manager.md @@ -8,4 +8,4 @@ especially useful for multi-speaker models. ```{eval-rst} .. automodule:: TTS.tts.utils.speakers :members: -``` \ No newline at end of file +``` diff --git a/docs/source/main_classes/trainer_api.md b/docs/source/main_classes/trainer_api.md index 876e09e5b6..bdb6048e45 100644 --- a/docs/source/main_classes/trainer_api.md +++ b/docs/source/main_classes/trainer_api.md @@ -1,3 +1,3 @@ # Trainer API -We made the trainer a separate project on https://github.com/coqui-ai/Trainer +We made the trainer a separate project: https://github.com/idiap/coqui-ai-Trainer diff --git a/docs/source/marytts.md b/docs/source/marytts.md index 9091ca330f..11cf4a2b9a 100644 --- a/docs/source/marytts.md +++ b/docs/source/marytts.md @@ -1,4 +1,4 @@ -# Mary-TTS API Support for Coqui-TTS +# Mary-TTS API support for Coqui TTS ## What is Mary-TTS? diff --git a/docs/source/models/bark.md b/docs/source/models/bark.md index c328ae6110..77f99c0d3a 100644 --- a/docs/source/models/bark.md +++ b/docs/source/models/bark.md @@ -37,7 +37,7 @@ from TTS.api import TTS # Load the model to GPU # Bark is really slow on CPU, so we recommend using GPU. -tts = TTS("tts_models/multilingual/multi-dataset/bark", gpu=True) +tts = TTS("tts_models/multilingual/multi-dataset/bark").to("cuda") # Cloning a new speaker @@ -57,7 +57,7 @@ tts.tts_to_file(text="Hello, my name is Manmay , how are you?", # random speaker -tts = TTS("tts_models/multilingual/multi-dataset/bark", gpu=True) +tts = TTS("tts_models/multilingual/multi-dataset/bark").to("cuda") tts.tts_to_file("hello world", file_path="out.wav") ``` @@ -69,14 +69,12 @@ tts --model_name tts_models/multilingual/multi-dataset/bark \ --text "This is an example." \ --out_path "output.wav" \ --voice_dir bark_voices/ \ ---speaker_idx "ljspeech" \ ---progress_bar True +--speaker_idx "ljspeech" # Random voice generation tts --model_name tts_models/multilingual/multi-dataset/bark \ --text "This is an example." \ ---out_path "output.wav" \ ---progress_bar True +--out_path "output.wav" ``` diff --git a/docs/source/models/forward_tts.md b/docs/source/models/forward_tts.md index f8f941c2fd..d618e4e056 100644 --- a/docs/source/models/forward_tts.md +++ b/docs/source/models/forward_tts.md @@ -61,5 +61,3 @@ Currently we provide the following pre-configured architectures: .. autoclass:: TTS.tts.configs.fast_speech_config.FastSpeechConfig :members: ``` - - diff --git a/docs/source/models/overflow.md b/docs/source/models/overflow.md index 09e270eae5..042ad47474 100644 --- a/docs/source/models/overflow.md +++ b/docs/source/models/overflow.md @@ -33,4 +33,4 @@ are available at https://shivammehta25.github.io/OverFlow/. ```{eval-rst} .. autoclass:: TTS.tts.models.overflow.Overflow :members: -``` \ No newline at end of file +``` diff --git a/docs/source/models/tacotron1-2.md b/docs/source/models/tacotron1-2.md index 25721eba4c..285d4f3c55 100644 --- a/docs/source/models/tacotron1-2.md +++ b/docs/source/models/tacotron1-2.md @@ -20,8 +20,8 @@ If you have a limited VRAM, then you can try using the Guided Attention Loss or ## Important resources & papers -- Tacotron: https://arxiv.org/abs/2006.06873 -- Tacotron2: https://arxiv.org/abs/2008.03802 +- Tacotron: [Tacotron: Towards End-to-End Speech Synthesis](https://arxiv.org/abs/1703.10135) +- Tacotron2: [Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions](https://arxiv.org/abs/1712.05884) - Double Decoder Consistency: https://coqui.ai/blog/tts/solving-attention-problems-of-tts-models-with-double-decoder-consistency - Guided Attention Loss: https://arxiv.org/abs/1710.08969 - Forward & Backward Decoder: https://arxiv.org/abs/1907.09006 @@ -59,5 +59,3 @@ If you have a limited VRAM, then you can try using the Guided Attention Loss or .. autoclass:: TTS.tts.configs.tacotron2_config.Tacotron2Config :members: ``` - - diff --git a/docs/source/models/tortoise.md b/docs/source/models/tortoise.md index 1a8e9ca8e9..30afd1355b 100644 --- a/docs/source/models/tortoise.md +++ b/docs/source/models/tortoise.md @@ -57,14 +57,12 @@ tts --model_name tts_models/en/multi-dataset/tortoise-v2 \ --text "This is an example." \ --out_path "output.wav" \ --voice_dir path/to/tortoise/voices/dir/ \ ---speaker_idx "lj" \ ---progress_bar True +--speaker_idx "lj" # Random voice generation tts --model_name tts_models/en/multi-dataset/tortoise-v2 \ --text "This is an example." \ ---out_path "output.wav" \ ---progress_bar True +--out_path "output.wav" ``` diff --git a/docs/source/models/xtts.md b/docs/source/models/xtts.md index b979d04f6e..91d4b4078c 100644 --- a/docs/source/models/xtts.md +++ b/docs/source/models/xtts.md @@ -1,59 +1,70 @@ -# ⓍTTS -ⓍTTS is a super cool Text-to-Speech model that lets you clone voices in different languages by using just a quick 3-second audio clip. Built on the đŸĸTortoise, -ⓍTTS has important model changes that make cross-language voice cloning and multi-lingual speech generation super easy. +# XTTS +XTTS is a super cool Text-to-Speech model that lets you clone voices in different languages by using just a quick 3-second audio clip. Built on the đŸĸTortoise, +XTTS has important model changes that make cross-language voice cloning and multi-lingual speech generation super easy. There is no need for an excessive amount of training data that spans countless hours. -This is the same model that powers [Coqui Studio](https://coqui.ai/), and [Coqui API](https://docs.coqui.ai/docs), however we apply -a few tricks to make it faster and support streaming inference. - -### Features +## Features - Voice cloning. - Cross-language voice cloning. - Multi-lingual speech generation. - 24khz sampling rate. -- Streaming inference with < 200ms latency. (See [Streaming inference](#streaming-inference)) +- Streaming inference with < 200ms latency. (See [Streaming inference](#streaming-manually)) - Fine-tuning support. (See [Training](#training)) -### Updates with v2 +## Updates with v2 - Improved voice cloning. - Voices can be cloned with a single audio file or multiple audio files, without any effect on the runtime. -- 2 new languages: Hungarian and Korean. - Across the board quality improvements. -### Code +## Code Current implementation only supports inference and GPT encoder training. -### Languages -As of now, XTTS-v2 supports 16 languages: English (en), Spanish (es), French (fr), German (de), Italian (it), Portuguese (pt), Polish (pl), Turkish (tr), Russian (ru), Dutch (nl), Czech (cs), Arabic (ar), Chinese (zh-cn), Japanese (ja), Hungarian (hu) and Korean (ko). - -Stay tuned as we continue to add support for more languages. If you have any language requests, please feel free to reach out. - -### License +## Languages +XTTS-v2 supports 17 languages: + +- Arabic (ar) +- Chinese (zh-cn) +- Czech (cs) +- Dutch (nl) +- English (en) +- French (fr) +- German (de) +- Hindi (hi) +- Hungarian (hu) +- Italian (it) +- Japanese (ja) +- Korean (ko) +- Polish (pl) +- Portuguese (pt) +- Russian (ru) +- Spanish (es) +- Turkish (tr) + +## License This model is licensed under [Coqui Public Model License](https://coqui.ai/cpml). -### Contact -Come and join in our 🐸Community. We're active on [Discord](https://discord.gg/fBC58unbKE) and [Twitter](https://twitter.com/coqui_ai). -You can also mail us at info@coqui.ai. +## Contact +Come and join in our 🐸Community. We're active on [Discord](https://discord.gg/fBC58unbKE) and [Github](https://github.com/idiap/coqui-ai-TTS/discussions). -### Inference +## Inference -#### 🐸TTS Command line +### 🐸TTS Command line -You can check all supported languages with the following command: +You can check all supported languages with the following command: ```console tts --model_name tts_models/multilingual/multi-dataset/xtts_v2 \ --list_language_idx ``` -You can check all Coqui available speakers with the following command: +You can check all Coqui available speakers with the following command: ```console tts --model_name tts_models/multilingual/multi-dataset/xtts_v2 \ --list_speaker_idx ``` -##### Coqui speakers +#### Coqui speakers You can do inference using one of the available speakers using the following command: ```console @@ -61,29 +72,29 @@ You can do inference using one of the available speakers using the following com --text "It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent." \ --speaker_idx "Ana Florence" \ --language_idx en \ - --use_cuda true + --use_cuda ``` -##### Clone a voice +#### Clone a voice You can clone a speaker voice using a single or multiple references: -###### Single reference +##### Single reference ```console tts --model_name tts_models/multilingual/multi-dataset/xtts_v2 \ --text "BugÃŧn okula gitmek istemiyorum." \ --speaker_wav /path/to/target/speaker.wav \ --language_idx tr \ - --use_cuda true + --use_cuda ``` -###### Multiple references +##### Multiple references ```console tts --model_name tts_models/multilingual/multi-dataset/xtts_v2 \ --text "BugÃŧn okula gitmek istemiyorum." \ --speaker_wav /path/to/target/speaker.wav /path/to/target/speaker_2.wav /path/to/target/speaker_3.wav \ --language_idx tr \ - --use_cuda true + --use_cuda ``` or for all wav files in a directory you can use: @@ -92,22 +103,22 @@ or for all wav files in a directory you can use: --text "BugÃŧn okula gitmek istemiyorum." \ --speaker_wav /path/to/target/*.wav \ --language_idx tr \ - --use_cuda true + --use_cuda ``` -#### 🐸TTS API +### 🐸TTS API -##### Clone a voice +#### Clone a voice You can clone a speaker voice using a single or multiple references: -###### Single reference +##### Single reference Splits the text into sentences and generates audio for each sentence. The audio files are then concatenated to produce the final audio. You can optionally disable sentence splitting for better coherence but more VRAM and possibly hitting models context length limit. ```python from TTS.api import TTS -tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2", gpu=True) +tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda") # generate speech by cloning a voice using default settings tts.tts_to_file(text="It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.", @@ -118,7 +129,7 @@ tts.tts_to_file(text="It took me quite a long time to develop a voice, and now t ) ``` -###### Multiple references +##### Multiple references You can pass multiple audio files to the `speaker_wav` argument for better voice cloning. @@ -126,15 +137,15 @@ You can pass multiple audio files to the `speaker_wav` argument for better voice from TTS.api import TTS # using the default version set in 🐸TTS -tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2", gpu=True) +tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda") # using a specific version # 👀 see the branch names for versions on https://huggingface.co/coqui/XTTS-v2/tree/main # ❗some versions might be incompatible with the API -tts = TTS("xtts_v2.0.2", gpu=True) +tts = TTS("xtts_v2.0.2").to("cuda") # getting the latest XTTS_v2 -tts = TTS("xtts", gpu=True) +tts = TTS("xtts").to("cuda") # generate speech by cloning a voice using default settings tts.tts_to_file(text="It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.", @@ -143,29 +154,30 @@ tts.tts_to_file(text="It took me quite a long time to develop a voice, and now t language="en") ``` -##### Coqui speakers +#### Coqui speakers You can do inference using one of the available speakers using the following code: ```python from TTS.api import TTS -tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2", gpu=True) +tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda") # generate speech by cloning a voice using default settings -tts.tts_to_file(text="It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.", - file_path="output.wav", - speaker="Ana Florence", - language="en", - split_sentences=True - ) +tts.tts_to_file( + text="It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.", + file_path="output.wav", + speaker="Ana Florence", + language="en", + split_sentences=True +) ``` -#### 🐸TTS Model API +### 🐸TTS Model API To use the model API, you need to download the model files and pass config and model file paths manually. -#### Manual Inference +### Manual Inference If you want to be able to `load_checkpoint` with `use_deepspeed=True` and **enjoy the speedup**, you need to install deepspeed first. @@ -173,7 +185,7 @@ If you want to be able to `load_checkpoint` with `use_deepspeed=True` and **enjo pip install deepspeed==0.10.3 ``` -##### inference parameters +#### Inference parameters - `text`: The text to be synthesized. - `language`: The language of the text to be synthesized. @@ -188,7 +200,7 @@ pip install deepspeed==0.10.3 - `enable_text_splitting`: Whether to split the text into sentences and generate audio for each sentence. It allows you to have infinite input length but might loose important context between sentences. Defaults to True. -##### Inference +#### Inference ```python @@ -219,8 +231,13 @@ out = model.inference( torchaudio.save("xtts.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000) ``` +You can also use the Coqui speakers: + +```python +gpt_cond_latent, speaker_embedding = model.speaker_manager.speakers["Ana Florence"].values() +``` -##### Streaming manually +#### Streaming manually Here the goal is to stream the audio as it is being generated. This is useful for real-time applications. Streaming inference is typically slower than regular inference, but it allows to get a first chunk of audio faster. @@ -264,9 +281,9 @@ torchaudio.save("xtts_streaming.wav", wav.squeeze().unsqueeze(0).cpu(), 24000) ``` -### Training +## Training -#### Easy training +### Easy training To make `XTTS_v2` GPT encoder training easier for beginner users we did a gradio demo that implements the whole fine-tuning pipeline. The gradio demo enables the user to easily do the following steps: - Preprocessing of the uploaded audio or audio files in 🐸 TTS coqui formatter @@ -275,12 +292,12 @@ To make `XTTS_v2` GPT encoder training easier for beginner users we did a gradio The user can run this gradio demo locally or remotely using a Colab Notebook. -##### Run demo on Colab +#### Run demo on Colab To make the `XTTS_v2` fine-tuning more accessible for users that do not have good GPUs available we did a Google Colab Notebook. The Colab Notebook is available [here](https://colab.research.google.com/drive/1GiI4_X724M8q2W-zZ-jXo7cWTV7RfaH-?usp=sharing). -To learn how to use this Colab Notebook please check the [XTTS fine-tuning video](). +To learn how to use this Colab Notebook please check the [XTTS fine-tuning video](https://www.youtube.com/watch?v=8tpDiiouGxc). If you are not able to acess the video you need to follow the steps: @@ -291,10 +308,10 @@ If you are not able to acess the video you need to follow the steps: 5. Soon the training is done you can go to the third Tab (3 - Inference) and then click on the button "Step 3 - Load Fine-tuned XTTS model" and wait until the fine-tuned model is loaded. Then you can do the inference on the model by clicking on the button "Step 4 - Inference". -##### Run demo locally +#### Run demo locally To run the demo locally you need to do the following steps: -1. Install 🐸 TTS following the instructions available [here](https://tts.readthedocs.io/en/dev/installation.html#installation). +1. Install 🐸 TTS following the instructions available [here](https://coqui-tts.readthedocs.io/en/latest/installation.html). 2. Install the Gradio demo requirements with the command `python3 -m pip install -r TTS/demos/xtts_ft_demo/requirements.txt` 3. Run the Gradio demo using the command `python3 TTS/demos/xtts_ft_demo/xtts_demo.py` 4. Follow the steps presented in the [tutorial video](https://www.youtube.com/watch?v=8tpDiiouGxc&feature=youtu.be) to be able to fine-tune and test the fine-tuned model. @@ -308,7 +325,7 @@ If you are not able to access the video, here is what you need to do: 4. Go to the third Tab (3 - Inference) and then click on the button "Step 3 - Load Fine-tuned XTTS model" and wait until the fine-tuned model is loaded. 5. Now you can run inference with the model by clicking on the button "Step 4 - Inference". -#### Advanced training +### Advanced training A recipe for `XTTS_v2` GPT encoder training using `LJSpeech` dataset is available at https://github.com/coqui-ai/TTS/tree/dev/recipes/ljspeech/xtts_v1/train_gpt_xtts.py @@ -382,6 +399,6 @@ torchaudio.save(OUTPUT_WAV_PATH, torch.tensor(out["wav"]).unsqueeze(0), 24000) ## XTTS Model ```{eval-rst} -.. autoclass:: TTS.tts.models.xtts.XTTS +.. autoclass:: TTS.tts.models.xtts.Xtts :members: ``` diff --git a/docs/source/project_structure.md b/docs/source/project_structure.md new file mode 100644 index 0000000000..af3e472adc --- /dev/null +++ b/docs/source/project_structure.md @@ -0,0 +1,30 @@ +# Project structure + +## Directory structure + +A non-comprehensive overview of the Coqui source code: + +| Directory | Contents | +| - | - | +| **Core** | | +| **[`TTS/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS)** | Main source code | +| **[`- .models.json`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS/.models.json)** | Pretrained model list | +| **[`- api.py`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS/api.py)** | Python API | +| **[`- bin/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS/bin)** | Executables and CLI | +| **[`- tts/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS/tts)** | Text-to-speech models | +| **[`- configs/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS/tts/configs)** | Model configurations | +| **[`- layers/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS/tts/layers)** | Model layer definitions | +| **[`- models/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS/tts/models)** | Model definitions | +| **[`- vc/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS/vc)** | Voice conversion models | +| `- (same)` | | +| **[`- vocoder/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS/vocoder)** | Vocoder models | +| `- (same)` | | +| **[`- encoder/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/TTS/encoder)** | Speaker encoder models | +| `- (same)` | | +| **Recipes/notebooks** | | +| **[`notebooks/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/notebooks)** | Jupyter Notebooks for model evaluation, parameter selection and data analysis | +| **[`recipes/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/recipes)** | Training recipes | +| **Others** | | +| **[`pyproject.toml`](https://github.com/idiap/coqui-ai-TTS/tree/dev/pyproject.toml)** | Project metadata, configuration and dependencies | +| **[`docs/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/docs)** | Documentation | +| **[`tests/`](https://github.com/idiap/coqui-ai-TTS/tree/dev/tests)** | Unit and integration tests | diff --git a/docs/source/server.md b/docs/source/server.md new file mode 100644 index 0000000000..3fa211d0d7 --- /dev/null +++ b/docs/source/server.md @@ -0,0 +1,30 @@ +# Demo server + +![server.gif](https://github.com/idiap/coqui-ai-TTS/raw/main/images/demo_server.gif) + +You can boot up a demo 🐸TTS server to run an inference with your models (make +sure to install the additional dependencies with `pip install coqui-tts[server]`). +Note that the server is not optimized for performance and does not support all +Coqui models yet. + +The demo server provides pretty much the same interface as the CLI command. + +```bash +tts-server -h # see the help +tts-server --list_models # list the available models. +``` + +Run a TTS model, from the release models list, with its default vocoder. +If the model you choose is a multi-speaker TTS model, you can select different speakers on the Web interface and synthesize +speech. + +```bash +tts-server --model_name "///" +``` + +Run a TTS and a vocoder model from the released model list. Note that not every vocoder is compatible with every TTS model. + +```bash +tts-server --model_name "///" \ + --vocoder_name "///" +``` diff --git a/docs/source/finetuning.md b/docs/source/training/finetuning.md similarity index 91% rename from docs/source/finetuning.md rename to docs/source/training/finetuning.md index 069f565137..fa2ed34a54 100644 --- a/docs/source/finetuning.md +++ b/docs/source/training/finetuning.md @@ -1,4 +1,4 @@ -# Fine-tuning a 🐸 TTS model +# Fine-tuning a model ## Fine-tuning @@ -21,17 +21,21 @@ them and fine-tune it for your own dataset. This will help you in two main ways: Fine-tuning comes to the rescue in this case. You can take one of our pre-trained models and fine-tune it on your own speech dataset and achieve reasonable results with only a couple of hours of data. - However, note that, fine-tuning does not ensure great results. The model performance still depends on the - {ref}`dataset quality ` and the hyper-parameters you choose for fine-tuning. Therefore, + However, note that, fine-tuning does not ensure great results. The model + performance still depends on the [dataset quality](../datasets/what_makes_a_good_dataset.md) + and the hyper-parameters you choose for fine-tuning. Therefore, it still takes a bit of tinkering. ## Steps to fine-tune a 🐸 TTS model +```{note} XTTS has separate fine-tuning scripts, see [here](../models/xtts.md#training). +``` + 1. Setup your dataset. You need to format your target dataset in a certain way so that 🐸TTS data loader will be able to load it for the - training. Please see {ref}`this page ` for more information about formatting. + training. Please see [this page](../datasets/formatting_your_dataset.md) for more information about formatting. 2. Choose the model you want to fine-tune. @@ -47,7 +51,8 @@ them and fine-tune it for your own dataset. This will help you in two main ways: You should choose the model based on your requirements. Some models are fast and some are better in speech quality. One lazy way to test a model is running the model on the hardware you want to use and see how it works. For - simple testing, you can use the `tts` command on the terminal. For more info see {ref}`here `. + simple testing, you can use the `tts` command on the terminal. For more info + see [here](../inference.md). 3. Download the model. @@ -111,4 +116,3 @@ them and fine-tune it for your own dataset. This will help you in two main ways: --coqpit.run_name "glow-tts-finetune" \ --coqpit.lr 0.00001 ``` - diff --git a/docs/source/training/index.md b/docs/source/training/index.md new file mode 100644 index 0000000000..b09f9cadcb --- /dev/null +++ b/docs/source/training/index.md @@ -0,0 +1,13 @@ +# Training and fine-tuning + +The following pages show you how to train and fine-tune Coqui models: + +```{toctree} +:maxdepth: 1 + +training_a_model +finetuning +``` + +Also see the [XTTS page](../models/xtts.md#training) if you want to fine-tune +that model. diff --git a/docs/source/training_a_model.md b/docs/source/training/training_a_model.md similarity index 92% rename from docs/source/training_a_model.md rename to docs/source/training/training_a_model.md index 989a57042a..22505ccb17 100644 --- a/docs/source/training_a_model.md +++ b/docs/source/training/training_a_model.md @@ -1,4 +1,4 @@ -# Training a Model +# Training a model 1. Decide the model you want to use. @@ -11,11 +11,10 @@ 3. Check the recipes. - Recipes are located under `TTS/recipes/`. They do not promise perfect models but they provide a good start point for - `Nervous Beginners`. + Recipes are located under `TTS/recipes/`. They do not promise perfect models but they provide a good start point. A recipe for `GlowTTS` using `LJSpeech` dataset looks like below. Let's be creative and call this `train_glowtts.py`. - ```{literalinclude} ../../recipes/ljspeech/glow_tts/train_glowtts.py + ```{literalinclude} ../../../recipes/ljspeech/glow_tts/train_glowtts.py ``` You need to change fields of the `BaseDatasetConfig` to match your dataset and then update `GlowTTSConfig` @@ -113,7 +112,7 @@ Note that different models have different metrics, visuals and outputs. - You should also check the [FAQ page](https://github.com/coqui-ai/TTS/wiki/FAQ) for common problems and solutions + You should also check the [FAQ page](../faq.md) for common problems and solutions that occur in a training. 7. Use your best model for inference. @@ -132,7 +131,7 @@ In the example above, we trained a `GlowTTS` model, but the same workflow applies to all the other 🐸TTS models. -# Multi-speaker Training +## Multi-speaker Training Training a multi-speaker model is mostly the same as training a single-speaker model. You need to specify a couple of configuration parameters, initiate a `SpeakerManager` instance and pass it to the model. @@ -142,5 +141,5 @@ d-vectors. For using d-vectors, you first need to compute the d-vectors using th The same Glow-TTS model above can be trained on a multi-speaker VCTK dataset with the script below. -```{literalinclude} ../../recipes/vctk/glow_tts/train_glow_tts.py +```{literalinclude} ../../../recipes/vctk/glow_tts/train_glow_tts.py ``` diff --git a/docs/source/tutorial_for_nervous_beginners.md b/docs/source/tutorial_for_nervous_beginners.md index acde3fc4c2..5e5eac0e0a 100644 --- a/docs/source/tutorial_for_nervous_beginners.md +++ b/docs/source/tutorial_for_nervous_beginners.md @@ -1,24 +1,40 @@ -# Tutorial For Nervous Beginners +# Tutorial for nervous beginners -## Installation +First [install](installation.md) Coqui TTS. -User friendly installation. Recommended only for synthesizing voice. +## Synthesizing Speech + +You can run `tts` and synthesize speech directly on the terminal. ```bash -$ pip install TTS +$ tts -h # see the help +$ tts --list_models # list the available models. ``` -Developer friendly installation. +![cli.gif](https://github.com/idiap/coqui-ai-TTS/raw/main/images/tts_cli.gif) + + +You can call `tts-server` to start a local demo server that you can open on +your favorite web browser and đŸ—Ŗī¸ (make sure to install the additional +dependencies with `pip install coqui-tts[server]`). ```bash -$ git clone https://github.com/coqui-ai/TTS -$ cd TTS -$ pip install -e . +$ tts-server -h # see the help +$ tts-server --list_models # list the available models. ``` +![server.gif](https://github.com/idiap/coqui-ai-TTS/raw/main/images/demo_server.gif) + +See [this page](inference.md) for more details on synthesizing speech with the +CLI, server or Python API. ## Training a `tts` Model -A breakdown of a simple script that trains a GlowTTS model on the LJspeech dataset. See the comments for more details. +```{note} XTTS has separate fine-tuning scripts, see [here](models/xtts.md#training). +``` + +A breakdown of a simple script that trains a GlowTTS model on the LJspeech +dataset. For a more in-depth guide to training and fine-tuning also see [this +page](training/index.md). ### Pure Python Way @@ -99,24 +115,3 @@ We still support running training from CLI like in the old days. The same traini ``` ❗ī¸ Note that you can also use ```train_vocoder.py``` as the ```tts``` models above. - -## Synthesizing Speech - -You can run `tts` and synthesize speech directly on the terminal. - -```bash -$ tts -h # see the help -$ tts --list_models # list the available models. -``` - -![cli.gif](https://github.com/coqui-ai/TTS/raw/main/images/tts_cli.gif) - - -You can call `tts-server` to start a local demo server that you can open it on -your favorite web browser and đŸ—Ŗī¸. - -```bash -$ tts-server -h # see the help -$ tts-server --list_models # list the available models. -``` -![server.gif](https://github.com/coqui-ai/TTS/raw/main/images/demo_server.gif) diff --git a/hubconf.py b/hubconf.py index 0c9c5930fc..6e10928265 100644 --- a/hubconf.py +++ b/hubconf.py @@ -1,15 +1,11 @@ -dependencies = [ - 'torch', 'gdown', 'pysbd', 'gruut', 'anyascii', 'pypinyin', 'coqpit', 'mecab-python3', 'unidic-lite' -] +dependencies = ["torch", "gdown", "pysbd", "gruut", "anyascii", "pypinyin", "coqpit", "mecab-python3", "unidic-lite"] import torch from TTS.utils.manage import ModelManager from TTS.utils.synthesizer import Synthesizer -def tts(model_name='tts_models/en/ljspeech/tacotron2-DCA', - vocoder_name=None, - use_cuda=False): +def tts(model_name="tts_models/en/ljspeech/tacotron2-DCA", vocoder_name=None, use_cuda=False): """TTS entry point for PyTorch Hub that provides a Synthesizer object to synthesize speech from a give text. Example: @@ -28,19 +24,20 @@ def tts(model_name='tts_models/en/ljspeech/tacotron2-DCA', manager = ModelManager() model_path, config_path, model_item = manager.download_model(model_name) - vocoder_name = model_item[ - 'default_vocoder'] if vocoder_name is None else vocoder_name + vocoder_name = model_item["default_vocoder"] if vocoder_name is None else vocoder_name vocoder_path, vocoder_config_path, _ = manager.download_model(vocoder_name) # create synthesizer - synt = Synthesizer(tts_checkpoint=model_path, - tts_config_path=config_path, - vocoder_checkpoint=vocoder_path, - vocoder_config=vocoder_config_path, - use_cuda=use_cuda) + synt = Synthesizer( + tts_checkpoint=model_path, + tts_config_path=config_path, + vocoder_checkpoint=vocoder_path, + vocoder_config=vocoder_config_path, + use_cuda=use_cuda, + ) return synt -if __name__ == '__main__': - synthesizer = torch.hub.load('coqui-ai/TTS:dev', 'tts', source='github') +if __name__ == "__main__": + synthesizer = torch.hub.load("coqui-ai/TTS:dev", "tts", source="github") synthesizer.tts("This is a test!") diff --git a/images/TTS-performance.png b/images/TTS-performance.png deleted file mode 100644 index 68eebaf7e6..0000000000 Binary files a/images/TTS-performance.png and /dev/null differ diff --git a/images/tts_performance.png b/images/tts_performance.png deleted file mode 100644 index bdff06731e..0000000000 Binary files a/images/tts_performance.png and /dev/null differ diff --git a/notebooks/TestAttention.ipynb b/notebooks/TestAttention.ipynb index 65edf98ca4..f52fa028e5 100644 --- a/notebooks/TestAttention.ipynb +++ b/notebooks/TestAttention.ipynb @@ -119,9 +119,9 @@ "\n", "# load model state\n", "if use_cuda:\n", - " cp = torch.load(MODEL_PATH)\n", + " cp = torch.load(MODEL_PATH, weights_only=True)\n", "else:\n", - " cp = torch.load(MODEL_PATH, map_location=lambda storage, loc: storage)\n", + " cp = torch.load(MODEL_PATH, map_location=lambda storage, loc: storage, weights_only=True)\n", "\n", "# load the model\n", "model.load_state_dict(cp['model'])\n", @@ -185,4 +185,4 @@ }, "nbformat": 4, "nbformat_minor": 4 -} \ No newline at end of file +} diff --git a/notebooks/Tutorial_1_use-pretrained-TTS.ipynb b/notebooks/Tutorial_1_use-pretrained-TTS.ipynb index 87d04c499d..3c2e9de924 100644 --- a/notebooks/Tutorial_1_use-pretrained-TTS.ipynb +++ b/notebooks/Tutorial_1_use-pretrained-TTS.ipynb @@ -41,7 +41,7 @@ "outputs": [], "source": [ "! pip install -U pip\n", - "! pip install TTS" + "! pip install coqui-tts" ] }, { diff --git a/notebooks/Tutorial_2_train_your_first_TTS_model.ipynb b/notebooks/Tutorial_2_train_your_first_TTS_model.ipynb index 0f580a85b6..c4186670c9 100644 --- a/notebooks/Tutorial_2_train_your_first_TTS_model.ipynb +++ b/notebooks/Tutorial_2_train_your_first_TTS_model.ipynb @@ -32,7 +32,7 @@ "source": [ "## Install Coqui TTS\n", "! pip install -U pip\n", - "! pip install TTS" + "! pip install coqui-tts" ] }, { @@ -44,7 +44,7 @@ "\n", "### **First things first**: we need some data.\n", "\n", - "We're training a Text-to-Speech model, so we need some _text_ and we need some _speech_. Specificially, we want _transcribed speech_. The speech must be divided into audio clips and each clip needs transcription. More details about data requirements such as recording characteristics, background noise and vocabulary coverage can be found in the [🐸TTS documentation](https://tts.readthedocs.io/en/latest/formatting_your_dataset.html).\n", + "We're training a Text-to-Speech model, so we need some _text_ and we need some _speech_. Specificially, we want _transcribed speech_. The speech must be divided into audio clips and each clip needs transcription. More details about data requirements such as recording characteristics, background noise and vocabulary coverage can be found in the [🐸TTS documentation](https://coqui-tts.readthedocs.io/en/latest/formatting_your_dataset.html).\n", "\n", "If you have a single audio file and you need to **split** it into clips. It is also important to use a lossless audio file format to prevent compression artifacts. We recommend using **wav** file format.\n", "\n", diff --git a/notebooks/dataset_analysis/CheckPitch.ipynb b/notebooks/dataset_analysis/CheckPitch.ipynb index 72afbc64a1..ebdac87378 100644 --- a/notebooks/dataset_analysis/CheckPitch.ipynb +++ b/notebooks/dataset_analysis/CheckPitch.ipynb @@ -176,4 +176,4 @@ }, "nbformat": 4, "nbformat_minor": 2 -} \ No newline at end of file +} diff --git a/notebooks/dataset_analysis/README.md b/notebooks/dataset_analysis/README.md index 79faf52159..9fe40d01a4 100644 --- a/notebooks/dataset_analysis/README.md +++ b/notebooks/dataset_analysis/README.md @@ -2,6 +2,6 @@ By the use of this notebook, you can easily analyze a brand new dataset, find exceptional cases and define your training set. -What we are looking in here is reasonable distribution of instances in terms of sequence-length, audio-length and word-coverage. +What we are looking in here is reasonable distribution of instances in terms of sequence-length, audio-length and word-coverage. This notebook is inspired from https://github.com/MycroftAI/mimic2 diff --git a/pyproject.toml b/pyproject.toml index 922575305c..a7baf29e31 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,20 +1,244 @@ +# ,*++++++*, ,*++++++*, +# *++. .+++ *++. .++* +# *+* ,++++* *+* *+* ,++++, *+* +# ,+, .++++++++++* ,++,,,,*+, ,++++++++++. *+, +# *+. .++++++++++++..++ *+.,++++++++++++. .+* +# .+* ++++++++++++.*+, .+*.++++++++++++ *+, +# .++ *++++++++* ++, .++.*++++++++* ++, +# ,+++*. . .*++, ,++*. .*+++* +# *+, .,*++**. .**++**. ,+* +# .+* *+, +# *+. Coqui .+* +# *+* +++ TTS +++ *+* +# .+++*. . . *+++. +# ,+* *+++*... ...*+++* *+, +# .++. .""""+++++++****+++++++"""". ++. +# ,++. .++, +# .++* *++. +# *+++, ,+++* +# .,*++++::::::++++*,. +# `````` + [build-system] -requires = [ - "setuptools", - "wheel", - "cython~=0.29.30", - "numpy>=1.22.0", - "packaging", +requires = ["hatchling"] +build-backend = "hatchling.build" + +[project] +name = "coqui-tts" +version = "0.25.1" +description = "Deep learning for Text to Speech." +readme = "README.md" +requires-python = ">=3.9, <3.13" +license = {text = "MPL-2.0"} +authors = [ + {name = "Eren GÃļlge", email = "egolge@coqui.ai"} +] +maintainers = [ + {name = "Enno Hermann", email = "enno.hermann@gmail.com"} +] +classifiers = [ + "Programming Language :: Python", + "Programming Language :: Python :: 3", + "Programming Language :: Python :: 3.9", + "Programming Language :: Python :: 3.10", + "Programming Language :: Python :: 3.11", + "Programming Language :: Python :: 3.12", + "Development Status :: 3 - Alpha", + "Intended Audience :: Science/Research", + "Intended Audience :: Developers", + "Operating System :: POSIX :: Linux", + "License :: OSI Approved :: Mozilla Public License 2.0 (MPL 2.0)", + "Topic :: Software Development", + "Topic :: Software Development :: Libraries :: Python Modules", + "Topic :: Multimedia :: Sound/Audio :: Speech", + "Topic :: Multimedia :: Sound/Audio", + "Topic :: Multimedia", + "Topic :: Scientific/Engineering :: Artificial Intelligence", +] +dependencies = [ + # Core + "numpy>=1.25.2,<2.0", + "cython>=3.0.0", + "scipy>=1.11.2", + "torch>=2.1", + "torchaudio>=2.1.0", + "soundfile>=0.12.0", + "librosa>=0.10.1", + "inflect>=5.6.0", + "tqdm>=4.64.1", + "anyascii>=0.3.0", + "pyyaml>=6.0", + "fsspec[http]>=2023.6.0", + "packaging>=23.1", + # Inference + "pysbd>=0.3.4", + # Training + "matplotlib>=3.7.0", + # Coqui stack + "coqui-tts-trainer>=0.2.0,<0.3.0", + "coqpit-config>=0.1.1,<0.2.0", + "monotonic-alignment-search>=0.1.0", + # Gruut + supported languages + "gruut[de,es,fr]>=2.4.0", + # Tortoise + "einops>=0.6.0", + "transformers>=4.43.0,<=4.46.2", + # Bark + "encodec>=0.1.1", + # XTTS + "num2words>=0.5.14", + "spacy[ja]>=3,<3.8", +] + +[project.optional-dependencies] +# Only used in notebooks +notebooks = [ + "bokeh==1.4.0", + "pandas>=1.4,<2.0", + "umap-learn>=0.5.1", +] +# For running the TTS server +server = ["flask>=3.0.0"] +# Language-specific dependencies, mainly for G2P +# Bangla +bn = [ + "bangla>=0.0.2", + "bnnumerizer>=0.0.2", + "bnunicodenormalizer>=0.1.0", +] +# Korean +ko = [ + "hangul_romanize>=0.1.0", + "jamo>=0.4.1", + "g2pkk>=0.1.1", + "pip>=22.2", +] +# Japanese +ja = [ + "mecab-python3>=1.0.2", + "unidic-lite==1.0.8", + "cutlet>=0.2.0", +] +# Chinese +zh = [ + "jieba>=0.42.1", + "pypinyin>=0.40.0", +] +# All language-specific dependencies +languages = [ + "coqui-tts[bn,ja,ko,zh]", +] +# Installs all extras (except dev and docs) +all = [ + "coqui-tts[notebooks,server,bn,ja,ko,zh]", +] + +[dependency-groups] +dev = [ + "black==24.2.0", + "coverage[toml]>=7", + "nose2>=0.15", + "pre-commit>=3", + "ruff==0.7.0", +] +# Dependencies for building the documentation +docs = [ + "furo>=2024.8.6", + "myst-parser==3.0.1", + "sphinx==7.4.7", + "sphinx_inline_tabs>=2023.4.21", + "sphinx_copybutton>=0.5.2", + "linkify-it-py>=2.0.3", ] -[flake8] -max-line-length=120 +[project.urls] +Homepage = "https://github.com/idiap/coqui-ai-TTS" +Documentation = "https://coqui-tts.readthedocs.io" +Repository = "https://github.com/idiap/coqui-ai-TTS" +Issues = "https://github.com/idiap/coqui-ai-TTS/issues" +Discussions = "https://github.com/idiap/coqui-ai-TTS/discussions" + +[project.scripts] +tts = "TTS.bin.synthesize:main" +tts-server = "TTS.server.server:main" + +[tool.uv] +constraint-dependencies = ["numba>0.58.0"] + +[tool.hatch.build] +exclude = [ + "/.github", + "/.gitignore", + "/.pre-commit-config.yaml", + "/.readthedocs.yml", + "/Makefile", + "/dockerfiles", + "/run_bash_tests.sh", + "/scripts", + "/tests", +] + +[tool.hatch.build.targets.wheel] +packages = ["TTS"] + +[tool.ruff] +line-length = 120 +extend-exclude = ["*.ipynb"] +lint.extend-select = [ + "B033", # duplicate-value + "C416", # unnecessary-comprehension + "D419", # empty-docstring + "F401", # unused-import + "F704", # yield-outside-function + "F706", # return-outside-function + "F841", # unused-variable + "I", # import sorting + "PIE790", # unnecessary-pass + "PLC", + "PLE", + "PLR0124", # comparison-with-itself + "PLR0206", # property-with-parameters + "PLR0911", # too-many-return-statements + "PLR1711", # useless-return + "PLW", + "W291", # trailing-whitespace + "NPY201", # NumPy 2.0 deprecation +] + +lint.ignore = [ + "E722", # bare except (TODO: fix these) + "E731", # don't use lambdas + "E741", # ambiguous variable name + "F821", # TODO: enable + "F841", # TODO: enable + "PLW0602", # TODO: enable + "PLW2901", # TODO: enable + "PLW0127", # TODO: enable + "PLW0603", # TODO: enable +] + +[tool.ruff.lint.pylint] +max-args = 5 +max-public-methods = 20 +max-returns = 7 + +[tool.ruff.lint.per-file-ignores] +"**/__init__.py" = [ + "F401", # init files may have "unused" imports for now + "F403", # init files may have star imports for now +] +"hubconf.py" = [ + "E402", # module level import not at top of file +] [tool.black] line-length = 120 target-version = ['py39'] -[tool.isort] -line_length = 120 -profile = "black" -multi_line_output = 3 +[tool.coverage.run] +parallel = true +source = ["TTS"] + +[tool.cibuildwheel] +build = "cp*" +skip = "*-win32 *i686 *musllinux*" diff --git a/recipes/README.md b/recipes/README.md index 21a6727d8b..fcc4719aaa 100644 --- a/recipes/README.md +++ b/recipes/README.md @@ -19,4 +19,4 @@ python TTS/bin/resample.py --input_dir recipes/vctk/VCTK/wav48_silence_trimmed - If you train a new model using TTS, feel free to share your training to expand the list of recipes. -You can also open a new discussion and share your progress with the 🐸 community. \ No newline at end of file +You can also open a new discussion and share your progress with the 🐸 community. diff --git a/recipes/bel-alex73/README.md b/recipes/bel-alex73/README.md index ad378dd998..6075d3102d 100644 --- a/recipes/bel-alex73/README.md +++ b/recipes/bel-alex73/README.md @@ -39,7 +39,7 @@ Docker container was created for simplify local running. You can run `docker-pre ## Training - with GPU -You need to upload Coqui-TTS(/mycomputer/TTS/) and storage/ directory(/mycomputer/storage/) to some computer with GPU. We don't need cv-corpus/ and fanetyka/ directories for training. Install gcc, then run `pip install -e .[all,dev,notebooks]` to prepare modules. GlowTTS and HifiGan models should be learned separately based on /storage/filtered_dataset only, i.e. they are not dependent from each other. below means list of GPU ids from zero("0,1,2,3" for systems with 4 GPU). See details on the https://tts.readthedocs.io/en/latest/tutorial_for_nervous_beginners.html(multi-gpu training). +You need to upload Coqui-TTS(/mycomputer/TTS/) and storage/ directory(/mycomputer/storage/) to some computer with GPU. We don't need cv-corpus/ and fanetyka/ directories for training. Install gcc, then run `pip install -e .[all,dev,notebooks]` to prepare modules. GlowTTS and HifiGan models should be learned separately based on /storage/filtered_dataset only, i.e. they are not dependent from each other. below means list of GPU ids from zero("0,1,2,3" for systems with 4 GPU). See details on the https://coqui-tts.readthedocs.io/en/latest/tutorial_for_nervous_beginners.html (multi-gpu training). Current setup created for 24GiB GPU. You need to change batch_size if you have more or less GPU memory. Also, you can try to set lr(learning rate) to lower value in the end of training GlowTTS. diff --git a/recipes/bel-alex73/train_hifigan.py b/recipes/bel-alex73/train_hifigan.py index 3e740b2ff4..78221a9f2b 100644 --- a/recipes/bel-alex73/train_hifigan.py +++ b/recipes/bel-alex73/train_hifigan.py @@ -1,11 +1,8 @@ -import os - -from coqpit import Coqpit from trainer import Trainer, TrainerArgs from TTS.tts.configs.shared_configs import BaseAudioConfig from TTS.utils.audio import AudioProcessor -from TTS.vocoder.configs.hifigan_config import * +from TTS.vocoder.configs.hifigan_config import HifiganConfig from TTS.vocoder.datasets.preprocess import load_wav_data from TTS.vocoder.models.gan import GAN diff --git a/recipes/blizzard2013/README.md b/recipes/blizzard2013/README.md index 9dcb739728..75f17a5513 100644 --- a/recipes/blizzard2013/README.md +++ b/recipes/blizzard2013/README.md @@ -9,4 +9,4 @@ To get a license and download link for this dataset, you need to visit the [webs You get access to the raw dataset in a couple of days. There are a few preprocessing steps you need to do to be able to use the high fidelity dataset. 1. Get the forced time alignments for the blizzard dataset from [here](https://github.com/mueller91/tts_alignments). -2. Segment the high fidelity audio-book files based on the instructions [here](https://github.com/Tomiinek/Blizzard2013_Segmentation). \ No newline at end of file +2. Segment the high fidelity audio-book files based on the instructions [here](https://github.com/Tomiinek/Blizzard2013_Segmentation). diff --git a/recipes/kokoro/tacotron2-DDC/run.sh b/recipes/kokoro/tacotron2-DDC/run.sh index 69800cf7b4..3f18f2c3fb 100644 --- a/recipes/kokoro/tacotron2-DDC/run.sh +++ b/recipes/kokoro/tacotron2-DDC/run.sh @@ -20,4 +20,4 @@ CUDA_VISIBLE_DEVICES="0" python TTS/bin/train_tts.py --config_path $RUN_DIR/taco --coqpit.output_path $RUN_DIR \ --coqpit.datasets.0.path $RUN_DIR/$CORPUS \ --coqpit.audio.stats_path $RUN_DIR/scale_stats.npy \ - --coqpit.phoneme_cache_path $RUN_DIR/phoneme_cache \ \ No newline at end of file + --coqpit.phoneme_cache_path $RUN_DIR/phoneme_cache \ diff --git a/recipes/kokoro/tacotron2-DDC/tacotron2-DDC.json b/recipes/kokoro/tacotron2-DDC/tacotron2-DDC.json index c2e526f46c..f422203a31 100644 --- a/recipes/kokoro/tacotron2-DDC/tacotron2-DDC.json +++ b/recipes/kokoro/tacotron2-DDC/tacotron2-DDC.json @@ -122,4 +122,4 @@ "use_gst": false, "use_external_speaker_embedding_file": false, "external_speaker_embedding_file": "../../speakers-vctk-en.json" -} \ No newline at end of file +} diff --git a/recipes/ljspeech/download_ljspeech.sh b/recipes/ljspeech/download_ljspeech.sh index 9468988a99..21c3e0e2d7 100644 --- a/recipes/ljspeech/download_ljspeech.sh +++ b/recipes/ljspeech/download_ljspeech.sh @@ -11,4 +11,4 @@ shuf LJSpeech-1.1/metadata.csv > LJSpeech-1.1/metadata_shuf.csv head -n 12000 LJSpeech-1.1/metadata_shuf.csv > LJSpeech-1.1/metadata_train.csv tail -n 1100 LJSpeech-1.1/metadata_shuf.csv > LJSpeech-1.1/metadata_val.csv mv LJSpeech-1.1 $RUN_DIR/recipes/ljspeech/ -rm LJSpeech-1.1.tar.bz2 \ No newline at end of file +rm LJSpeech-1.1.tar.bz2 diff --git a/recipes/ljspeech/fast_pitch/train_fast_pitch.py b/recipes/ljspeech/fast_pitch/train_fast_pitch.py index 055526b1bc..64fd737b4e 100644 --- a/recipes/ljspeech/fast_pitch/train_fast_pitch.py +++ b/recipes/ljspeech/fast_pitch/train_fast_pitch.py @@ -65,7 +65,7 @@ model_path, config_path, _ = manager.download_model("tts_models/en/ljspeech/tacotron2-DCA") # TODO: make compute_attention python callable os.system( - f"python TTS/bin/compute_attention_masks.py --model_path {model_path} --config_path {config_path} --dataset ljspeech --dataset_metafile metadata.csv --data_path ./recipes/ljspeech/LJSpeech-1.1/ --use_cuda true" + f"python TTS/bin/compute_attention_masks.py --model_path {model_path} --config_path {config_path} --dataset ljspeech --dataset_metafile metadata.csv --data_path ./recipes/ljspeech/LJSpeech-1.1/ --use_cuda" ) # INITIALIZE THE AUDIO PROCESSOR diff --git a/recipes/ljspeech/fast_speech/train_fast_speech.py b/recipes/ljspeech/fast_speech/train_fast_speech.py index 8c9a272e81..9839fcb339 100644 --- a/recipes/ljspeech/fast_speech/train_fast_speech.py +++ b/recipes/ljspeech/fast_speech/train_fast_speech.py @@ -64,7 +64,7 @@ model_path, config_path, _ = manager.download_model("tts_models/en/ljspeech/tacotron2-DCA") # TODO: make compute_attention python callable os.system( - f"python TTS/bin/compute_attention_masks.py --model_path {model_path} --config_path {config_path} --dataset ljspeech --dataset_metafile metadata.csv --data_path ./recipes/ljspeech/LJSpeech-1.1/ --use_cuda true" + f"python TTS/bin/compute_attention_masks.py --model_path {model_path} --config_path {config_path} --dataset ljspeech --dataset_metafile metadata.csv --data_path ./recipes/ljspeech/LJSpeech-1.1/ --use_cuda" ) # INITIALIZE THE AUDIO PROCESSOR diff --git a/recipes/ljspeech/fastspeech2/train_fastspeech2.py b/recipes/ljspeech/fastspeech2/train_fastspeech2.py index 93737dba7f..0a7a175605 100644 --- a/recipes/ljspeech/fastspeech2/train_fastspeech2.py +++ b/recipes/ljspeech/fastspeech2/train_fastspeech2.py @@ -67,7 +67,7 @@ model_path, config_path, _ = manager.download_model("tts_models/en/ljspeech/tacotron2-DCA") # TODO: make compute_attention python callable os.system( - f"python TTS/bin/compute_attention_masks.py --model_path {model_path} --config_path {config_path} --dataset ljspeech --dataset_metafile metadata.csv --data_path ./recipes/ljspeech/LJSpeech-1.1/ --use_cuda true" + f"python TTS/bin/compute_attention_masks.py --model_path {model_path} --config_path {config_path} --dataset ljspeech --dataset_metafile metadata.csv --data_path ./recipes/ljspeech/LJSpeech-1.1/ --use_cuda" ) # INITIALIZE THE AUDIO PROCESSOR diff --git a/recipes/ljspeech/xtts_v1/train_gpt_xtts.py b/recipes/ljspeech/xtts_v1/train_gpt_xtts.py index 7d8f4064c5..a077a18064 100644 --- a/recipes/ljspeech/xtts_v1/train_gpt_xtts.py +++ b/recipes/ljspeech/xtts_v1/train_gpt_xtts.py @@ -4,7 +4,8 @@ from TTS.config.shared_configs import BaseDatasetConfig from TTS.tts.datasets import load_tts_samples -from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig, XttsAudioConfig +from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig +from TTS.tts.models.xtts import XttsAudioConfig from TTS.utils.manage import ModelManager # Logging parameters @@ -41,8 +42,8 @@ # DVAE files -DVAE_CHECKPOINT_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v1/v1.1.2/dvae.pth" -MEL_NORM_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v1/v1.1.2/mel_stats.pth" +DVAE_CHECKPOINT_LINK = "https://huggingface.co/coqui/XTTS-v1/resolve/v1.1.2/dvae.pth" +MEL_NORM_LINK = "https://huggingface.co/coqui/XTTS-v1/resolve/v1.1.2/mel_stats.pth" # Set the path to the downloaded files DVAE_CHECKPOINT = os.path.join(CHECKPOINTS_OUT_PATH, DVAE_CHECKPOINT_LINK.split("/")[-1]) @@ -55,8 +56,8 @@ # Download XTTS v1.1 checkpoint if needed -TOKENIZER_FILE_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v1/v1.1.2/vocab.json" -XTTS_CHECKPOINT_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v1/v1.1.2/model.pth" +TOKENIZER_FILE_LINK = "https://huggingface.co/coqui/XTTS-v1/resolve/v1.1.2/vocab.json" +XTTS_CHECKPOINT_LINK = "https://huggingface.co/coqui/XTTS-v1/resolve/v1.1.2/model.pth" # XTTS transfer learning parameters: You we need to provide the paths of XTTS model checkpoint that you want to do the fine tuning. TOKENIZER_FILE = os.path.join(CHECKPOINTS_OUT_PATH, TOKENIZER_FILE_LINK.split("/")[-1]) # vocab.json file diff --git a/recipes/ljspeech/xtts_v2/train_gpt_xtts.py b/recipes/ljspeech/xtts_v2/train_gpt_xtts.py index 626917381a..362f45008e 100644 --- a/recipes/ljspeech/xtts_v2/train_gpt_xtts.py +++ b/recipes/ljspeech/xtts_v2/train_gpt_xtts.py @@ -4,7 +4,8 @@ from TTS.config.shared_configs import BaseDatasetConfig from TTS.tts.datasets import load_tts_samples -from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig, XttsAudioConfig +from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig +from TTS.tts.models.xtts import XttsAudioConfig from TTS.utils.manage import ModelManager # Logging parameters @@ -41,8 +42,8 @@ # DVAE files -DVAE_CHECKPOINT_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/dvae.pth" -MEL_NORM_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/mel_stats.pth" +DVAE_CHECKPOINT_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/dvae.pth" +MEL_NORM_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/mel_stats.pth" # Set the path to the downloaded files DVAE_CHECKPOINT = os.path.join(CHECKPOINTS_OUT_PATH, os.path.basename(DVAE_CHECKPOINT_LINK)) @@ -55,8 +56,8 @@ # Download XTTS v2.0 checkpoint if needed -TOKENIZER_FILE_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/vocab.json" -XTTS_CHECKPOINT_LINK = "https://coqui.gateway.scarf.sh/hf-coqui/XTTS-v2/main/model.pth" +TOKENIZER_FILE_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/vocab.json" +XTTS_CHECKPOINT_LINK = "https://huggingface.co/coqui/XTTS-v2/resolve/main/model.pth" # XTTS transfer learning parameters: You we need to provide the paths of XTTS model checkpoint that you want to do the fine tuning. TOKENIZER_FILE = os.path.join(CHECKPOINTS_OUT_PATH, os.path.basename(TOKENIZER_FILE_LINK)) # vocab.json file diff --git a/recipes/multilingual/cml_yourtts/train_yourtts.py b/recipes/multilingual/cml_yourtts/train_yourtts.py index 25a2fd0a4b..02f901fe73 100644 --- a/recipes/multilingual/cml_yourtts/train_yourtts.py +++ b/recipes/multilingual/cml_yourtts/train_yourtts.py @@ -4,7 +4,6 @@ from trainer import Trainer, TrainerArgs from TTS.bin.compute_embeddings import compute_embeddings -from TTS.bin.resample import resample_files from TTS.config.shared_configs import BaseDatasetConfig from TTS.tts.configs.vits_config import VitsConfig from TTS.tts.datasets import load_tts_samples diff --git a/recipes/thorsten_DE/align_tts/train_aligntts.py b/recipes/thorsten_DE/align_tts/train_aligntts.py index 32cfd9967f..42363940f3 100644 --- a/recipes/thorsten_DE/align_tts/train_aligntts.py +++ b/recipes/thorsten_DE/align_tts/train_aligntts.py @@ -30,7 +30,7 @@ run_eval=True, test_delay_epochs=-1, epochs=1000, - text_cleaner="phoneme_cleaners", + text_cleaner="multilingual_phoneme_cleaners", use_phonemes=False, phoneme_language="de", phoneme_cache_path=os.path.join(output_path, "phoneme_cache"), diff --git a/recipes/thorsten_DE/glow_tts/train_glowtts.py b/recipes/thorsten_DE/glow_tts/train_glowtts.py index 00c67fb5d8..f7f4a186a2 100644 --- a/recipes/thorsten_DE/glow_tts/train_glowtts.py +++ b/recipes/thorsten_DE/glow_tts/train_glowtts.py @@ -40,7 +40,7 @@ run_eval=True, test_delay_epochs=-1, epochs=1000, - text_cleaner="phoneme_cleaners", + text_cleaner="multilingual_phoneme_cleaners", use_phonemes=True, phoneme_language="de", phoneme_cache_path=os.path.join(output_path, "phoneme_cache"), diff --git a/recipes/thorsten_DE/speedy_speech/train_speedy_speech.py b/recipes/thorsten_DE/speedy_speech/train_speedy_speech.py index a3d0b9db2b..024dcaa31e 100644 --- a/recipes/thorsten_DE/speedy_speech/train_speedy_speech.py +++ b/recipes/thorsten_DE/speedy_speech/train_speedy_speech.py @@ -45,7 +45,7 @@ test_delay_epochs=-1, epochs=1000, min_audio_len=11050, # need to up min_audio_len to avois speedy speech error - text_cleaner="phoneme_cleaners", + text_cleaner="multilingual_phoneme_cleaners", use_phonemes=True, phoneme_language="de", phoneme_cache_path=os.path.join(output_path, "phoneme_cache"), diff --git a/recipes/thorsten_DE/tacotron2-DDC/train_tacotron_ddc.py b/recipes/thorsten_DE/tacotron2-DDC/train_tacotron_ddc.py index bc0274f5af..a46e27e91b 100644 --- a/recipes/thorsten_DE/tacotron2-DDC/train_tacotron_ddc.py +++ b/recipes/thorsten_DE/tacotron2-DDC/train_tacotron_ddc.py @@ -49,7 +49,7 @@ gradual_training=[[0, 6, 64], [10000, 4, 32], [50000, 3, 32], [100000, 2, 32]], double_decoder_consistency=True, epochs=1000, - text_cleaner="phoneme_cleaners", + text_cleaner="multilingual_phoneme_cleaners", use_phonemes=True, phoneme_language="de", phoneme_cache_path=os.path.join(output_path, "phoneme_cache"), diff --git a/recipes/thorsten_DE/vits_tts/train_vits.py b/recipes/thorsten_DE/vits_tts/train_vits.py index 4ffa0f30f6..4b773c3508 100644 --- a/recipes/thorsten_DE/vits_tts/train_vits.py +++ b/recipes/thorsten_DE/vits_tts/train_vits.py @@ -40,7 +40,7 @@ run_eval=True, test_delay_epochs=-1, epochs=1000, - text_cleaner="phoneme_cleaners", + text_cleaner="multilingual_phoneme_cleaners", use_phonemes=True, phoneme_language="de", phoneme_cache_path=os.path.join(output_path, "phoneme_cache"), diff --git a/requirements.dev.txt b/requirements.dev.txt deleted file mode 100644 index 8c674727d3..0000000000 --- a/requirements.dev.txt +++ /dev/null @@ -1,5 +0,0 @@ -black -coverage -isort -nose2 -pylint==2.10.2 diff --git a/requirements.ja.txt b/requirements.ja.txt deleted file mode 100644 index 4baab88a91..0000000000 --- a/requirements.ja.txt +++ /dev/null @@ -1,5 +0,0 @@ -# These cause some compatibility issues on some systems and are not strictly necessary -# japanese g2p deps -mecab-python3==1.0.6 -unidic-lite==1.0.8 -cutlet diff --git a/requirements.notebooks.txt b/requirements.notebooks.txt deleted file mode 100644 index 65d3f642c9..0000000000 --- a/requirements.notebooks.txt +++ /dev/null @@ -1 +0,0 @@ -bokeh==1.4.0 \ No newline at end of file diff --git a/requirements.txt b/requirements.txt deleted file mode 100644 index 2944e6face..0000000000 --- a/requirements.txt +++ /dev/null @@ -1,57 +0,0 @@ -# core deps -numpy==1.22.0;python_version<="3.10" -numpy>=1.24.3;python_version>"3.10" -cython>=0.29.30 -scipy>=1.11.2 -torch>=2.1 -torchaudio -soundfile>=0.12.0 -librosa>=0.10.0 -scikit-learn>=1.3.0 -numba==0.55.1;python_version<"3.9" -numba>=0.57.0;python_version>="3.9" -inflect>=5.6.0 -tqdm>=4.64.1 -anyascii>=0.3.0 -pyyaml>=6.0 -fsspec>=2023.6.0 # <= 2023.9.1 makes aux tests fail -aiohttp>=3.8.1 -packaging>=23.1 -mutagen==1.47.0 -# deps for examples -flask>=2.0.1 -# deps for inference -pysbd>=0.3.4 -# deps for notebooks -umap-learn>=0.5.1 -pandas>=1.4,<2.0 -# deps for training -matplotlib>=3.7.0 -# coqui stack -trainer>=0.0.36 -# config management -coqpit>=0.0.16 -# chinese g2p deps -jieba -pypinyin -# korean -hangul_romanize -# gruut+supported langs -gruut[de,es,fr]==2.2.3 -# deps for korean -jamo -nltk -g2pkk>=0.1.1 -# deps for bangla -bangla -bnnumerizer -bnunicodenormalizer -#deps for tortoise -einops>=0.6.0 -transformers>=4.33.0 -#deps for bark -encodec>=0.1.1 -# deps for XTTS -unidecode>=1.3.2 -num2words -spacy[ja]>=3 \ No newline at end of file diff --git a/scripts/sync_readme.py b/scripts/sync_readme.py index 584286814b..97256bca6d 100644 --- a/scripts/sync_readme.py +++ b/scripts/sync_readme.py @@ -22,8 +22,12 @@ def sync_readme(): new_content = replace_between_markers(orig_content, "tts-readme", description.strip()) if args.check: if orig_content != new_content: - print("README.md is out of sync; please edit TTS/bin/TTS_README.md and run scripts/sync_readme.py") + print( + "README.md is out of sync; please reconcile README.md and TTS/bin/synthesize.py and run scripts/sync_readme.py" + ) exit(42) + print("All good, files in sync") + exit(0) readme_path.write_text(new_content) print("Updated README.md") diff --git a/setup.cfg b/setup.cfg deleted file mode 100644 index 1f31cb5dec..0000000000 --- a/setup.cfg +++ /dev/null @@ -1,8 +0,0 @@ -[build_py] -build_lib=temp_build - -[bdist_wheel] -bdist_dir=temp_build - -[install_lib] -build_dir=temp_build diff --git a/setup.py b/setup.py deleted file mode 100644 index df14b41adc..0000000000 --- a/setup.py +++ /dev/null @@ -1,141 +0,0 @@ -#!/usr/bin/env python -# ,*++++++*, ,*++++++*, -# *++. .+++ *++. .++* -# *+* ,++++* *+* *+* ,++++, *+* -# ,+, .++++++++++* ,++,,,,*+, ,++++++++++. *+, -# *+. .++++++++++++..++ *+.,++++++++++++. .+* -# .+* ++++++++++++.*+, .+*.++++++++++++ *+, -# .++ *++++++++* ++, .++.*++++++++* ++, -# ,+++*. . .*++, ,++*. .*+++* -# *+, .,*++**. .**++**. ,+* -# .+* *+, -# *+. Coqui .+* -# *+* +++ TTS +++ *+* -# .+++*. . . *+++. -# ,+* *+++*... ...*+++* *+, -# .++. .""""+++++++****+++++++"""". ++. -# ,++. .++, -# .++* *++. -# *+++, ,+++* -# .,*++++::::::++++*,. -# `````` - -import os -import subprocess -import sys -from packaging.version import Version - -import numpy -import setuptools.command.build_py -import setuptools.command.develop -from Cython.Build import cythonize -from setuptools import Extension, find_packages, setup - -python_version = sys.version.split()[0] -if Version(python_version) < Version("3.9") or Version(python_version) >= Version("3.12"): - raise RuntimeError("TTS requires python >= 3.9 and < 3.12 " "but your Python version is {}".format(sys.version)) - - -cwd = os.path.dirname(os.path.abspath(__file__)) -with open(os.path.join(cwd, "TTS", "VERSION")) as fin: - version = fin.read().strip() - - -class build_py(setuptools.command.build_py.build_py): # pylint: disable=too-many-ancestors - def run(self): - setuptools.command.build_py.build_py.run(self) - - -class develop(setuptools.command.develop.develop): - def run(self): - setuptools.command.develop.develop.run(self) - - -# The documentation for this feature is in server/README.md -package_data = ["TTS/server/templates/*"] - - -def pip_install(package_name): - subprocess.call([sys.executable, "-m", "pip", "install", package_name]) - - -requirements = open(os.path.join(cwd, "requirements.txt"), "r").readlines() -with open(os.path.join(cwd, "requirements.notebooks.txt"), "r") as f: - requirements_notebooks = f.readlines() -with open(os.path.join(cwd, "requirements.dev.txt"), "r") as f: - requirements_dev = f.readlines() -with open(os.path.join(cwd, "requirements.ja.txt"), "r") as f: - requirements_ja = f.readlines() -requirements_all = requirements_dev + requirements_notebooks + requirements_ja - -with open("README.md", "r", encoding="utf-8") as readme_file: - README = readme_file.read() - -exts = [ - Extension( - name="TTS.tts.utils.monotonic_align.core", - sources=["TTS/tts/utils/monotonic_align/core.pyx"], - ) -] -setup( - name="TTS", - version=version, - url="https://github.com/coqui-ai/TTS", - author="Eren GÃļlge", - author_email="egolge@coqui.ai", - description="Deep learning for Text to Speech by Coqui.", - long_description=README, - long_description_content_type="text/markdown", - license="MPL-2.0", - # cython - include_dirs=numpy.get_include(), - ext_modules=cythonize(exts, language_level=3), - # ext_modules=find_cython_extensions(), - # package - include_package_data=True, - packages=find_packages(include=["TTS"], exclude=["*.tests", "*tests.*", "tests.*", "*tests", "tests"]), - package_data={ - "TTS": [ - "VERSION", - ] - }, - project_urls={ - "Documentation": "https://github.com/coqui-ai/TTS/wiki", - "Tracker": "https://github.com/coqui-ai/TTS/issues", - "Repository": "https://github.com/coqui-ai/TTS", - "Discussions": "https://github.com/coqui-ai/TTS/discussions", - }, - cmdclass={ - "build_py": build_py, - "develop": develop, - # 'build_ext': build_ext - }, - install_requires=requirements, - extras_require={ - "all": requirements_all, - "dev": requirements_dev, - "notebooks": requirements_notebooks, - "ja": requirements_ja, - }, - python_requires=">=3.9.0, <3.12", - entry_points={"console_scripts": ["tts=TTS.bin.synthesize:main", "tts-server = TTS.server.server:main"]}, - classifiers=[ - "Programming Language :: Python", - "Programming Language :: Python :: 3", - "Programming Language :: Python :: 3.9", - "Programming Language :: Python :: 3.10", - "Programming Language :: Python :: 3.11", - "Development Status :: 3 - Alpha", - "Intended Audience :: Science/Research", - "Intended Audience :: Developers", - "Operating System :: POSIX :: Linux", - "License :: OSI Approved :: Mozilla Public License 2.0 (MPL 2.0)", - "Topic :: Software Development", - "Topic :: Software Development :: Libraries :: Python Modules", - "Topic :: Multimedia :: Sound/Audio :: Speech", - "Topic :: Multimedia :: Sound/Audio", - "Topic :: Multimedia", - "Topic :: Scientific/Engineering :: Artificial Intelligence", - ], - zip_safe=False, -) diff --git a/tests/__init__.py b/tests/__init__.py index e102a2dfee..f0a8b2f118 100644 --- a/tests/__init__.py +++ b/tests/__init__.py @@ -1,7 +1,8 @@ import os +from trainer.generic_utils import get_cuda + from TTS.config import BaseDatasetConfig -from TTS.utils.generic_utils import get_cuda def get_device_id(): diff --git a/tests/tts_tests/test_helpers.py b/tests/aux_tests/test_helpers.py similarity index 67% rename from tests/tts_tests/test_helpers.py rename to tests/aux_tests/test_helpers.py index 23bb440a0a..6781cbc5d4 100644 --- a/tests/tts_tests/test_helpers.py +++ b/tests/aux_tests/test_helpers.py @@ -1,9 +1,17 @@ import torch as T -from TTS.tts.utils.helpers import average_over_durations, generate_path, rand_segments, segment, sequence_mask - - -def average_over_durations_test(): # pylint: disable=no-self-use +from TTS.tts.utils.helpers import ( + average_over_durations, + expand_encoder_outputs, + generate_attention, + generate_path, + rand_segments, + segment, + sequence_mask, +) + + +def test_average_over_durations(): # pylint: disable=no-self-use pitch = T.rand(1, 1, 128) durations = T.randint(1, 5, (1, 21)) @@ -21,7 +29,7 @@ def average_over_durations_test(): # pylint: disable=no-self-use index += dur -def seqeunce_mask_test(): +def test_sequence_mask(): lengths = T.randint(10, 15, (8,)) mask = sequence_mask(lengths) for i in range(8): @@ -30,8 +38,8 @@ def seqeunce_mask_test(): assert mask[i, l:].sum() == 0 -def segment_test(): - x = T.range(0, 11) +def test_segment(): + x = T.arange(0, 12) x = x.repeat(8, 1).unsqueeze(1) segment_ids = T.randint(0, 7, (8,)) @@ -50,11 +58,11 @@ def segment_test(): assert x[idx, :, start_indx : start_indx + 10].sum() == segments[idx, :, :].sum() -def rand_segments_test(): +def test_rand_segments(): x = T.rand(2, 3, 4) x_lens = T.randint(3, 4, (2,)) - segments, seg_idxs = rand_segments(x, x_lens, segment_size=3) - assert segments.shape == (2, 3, 3) + segments, seg_idxs = rand_segments(x, x_lens, segment_size=2) + assert segments.shape == (2, 3, 2) assert all(seg_idxs >= 0), seg_idxs try: segments, _ = rand_segments(x, x_lens, segment_size=5) @@ -68,10 +76,10 @@ def rand_segments_test(): assert all(x_lens_back == x_lens) -def generate_path_test(): +def test_generate_path(): durations = T.randint(1, 4, (10, 21)) x_length = T.randint(18, 22, (10,)) - x_mask = sequence_mask(x_length).unsqueeze(1).long() + x_mask = sequence_mask(x_length, max_len=21).unsqueeze(1).long() durations = durations * x_mask.squeeze(1) y_length = durations.sum(1) y_mask = sequence_mask(y_length).unsqueeze(1).long() @@ -86,3 +94,24 @@ def generate_path_test(): assert all(path[b, t, :current_idx] == 0.0) assert all(path[b, t, current_idx + durations[b, t].item() :] == 0.0) current_idx += durations[b, t].item() + + assert T.all(path == generate_attention(durations, x_mask, y_mask)) + assert T.all(path == generate_attention(durations, x_mask)) + + +def test_expand_encoder_outputs(): + inputs = T.rand(2, 5, 57) + durations = T.randint(1, 4, (2, 57)) + + x_mask = T.ones(2, 1, 57) + y_lengths = T.ones(2) * durations.sum(1).max() + + expanded, _, _ = expand_encoder_outputs(inputs, durations, x_mask, y_lengths) + + for b in range(durations.shape[0]): + index = 0 + for idx, dur in enumerate(durations[b]): + idx_expanded = expanded[b, :, index : index + dur.item()] + diff = (idx_expanded - inputs[b, :, idx].repeat(int(dur)).view(idx_expanded.shape)).sum() + assert abs(diff) < 1e-6, diff + index += dur diff --git a/tests/aux_tests/test_stft_torch.py b/tests/aux_tests/test_stft_torch.py deleted file mode 100644 index e69de29bb2..0000000000 diff --git a/tests/aux_tests/test_torch_transforms.py b/tests/aux_tests/test_torch_transforms.py new file mode 100644 index 0000000000..2da5a359c1 --- /dev/null +++ b/tests/aux_tests/test_torch_transforms.py @@ -0,0 +1,16 @@ +import numpy as np +import torch + +from TTS.utils.audio import numpy_transforms as np_transforms +from TTS.utils.audio.torch_transforms import amp_to_db, db_to_amp + + +def test_amplitude_db_conversion(): + x = torch.rand(11) + o1 = amp_to_db(x=x, spec_gain=1.0) + o2 = db_to_amp(x=o1, spec_gain=1.0) + np_o1 = np_transforms.amp_to_db(x=x, base=np.e) + np_o2 = np_transforms.db_to_amp(x=np_o1, base=np.e) + assert torch.allclose(x, o2) + assert torch.allclose(o1, np_o1) + assert torch.allclose(o2, np_o2) diff --git a/tests/bash_tests/test_compute_statistics.sh b/tests/bash_tests/test_compute_statistics.sh index d7f0ab9d4c..721777f852 100755 --- a/tests/bash_tests/test_compute_statistics.sh +++ b/tests/bash_tests/test_compute_statistics.sh @@ -4,4 +4,3 @@ BASEDIR=$(dirname "$0") echo "$BASEDIR" # run training CUDA_VISIBLE_DEVICES="" python TTS/bin/compute_statistics.py --config_path $BASEDIR/../inputs/test_glow_tts.json --out_path $BASEDIR/../outputs/scale_stats.npy - diff --git a/tests/data/dummy_speakers.json b/tests/data/dummy_speakers.json index 233533b796..507b57b5a5 100644 --- a/tests/data/dummy_speakers.json +++ b/tests/data/dummy_speakers.json @@ -100222,5 +100222,5 @@ 0.04999300092458725, -0.12125937640666962 ] - } + } } diff --git a/tests/data/ljspeech/metadata_flac.csv b/tests/data/ljspeech/metadata_flac.csv index 43db05ac91..fbde71d07d 100644 --- a/tests/data/ljspeech/metadata_flac.csv +++ b/tests/data/ljspeech/metadata_flac.csv @@ -6,4 +6,4 @@ wavs/LJ001-0004.flac|produced the block books, which were the immediate predeces wavs/LJ001-0005.flac|the invention of movable metal letters in the middle of the fifteenth century may justly be considered as the invention of the art of printing.|the invention of movable metal letters in the middle of the fifteenth century may justly be considered as the invention of the art of printing.|ljspeech-2 wavs/LJ001-0006.flac|And it is worth mention in passing that, as an example of fine typography,|And it is worth mention in passing that, as an example of fine typography,|ljspeech-2 wavs/LJ001-0007.flac|the earliest book printed with movable types, the Gutenberg, or "forty-two line Bible" of about 1455,|the earliest book printed with movable types, the Gutenberg, or "forty-two line Bible" of about fourteen fifty-five,|ljspeech-3 -wavs/LJ001-0008.flac|has never been surpassed.|has never been surpassed.|ljspeech-3 \ No newline at end of file +wavs/LJ001-0008.flac|has never been surpassed.|has never been surpassed.|ljspeech-3 diff --git a/tests/data/ljspeech/metadata_mp3.csv b/tests/data/ljspeech/metadata_mp3.csv index 109e48b40a..a8c5ec2e76 100644 --- a/tests/data/ljspeech/metadata_mp3.csv +++ b/tests/data/ljspeech/metadata_mp3.csv @@ -6,4 +6,4 @@ wavs/LJ001-0004.mp3|produced the block books, which were the immediate predecess wavs/LJ001-0005.mp3|the invention of movable metal letters in the middle of the fifteenth century may justly be considered as the invention of the art of printing.|the invention of movable metal letters in the middle of the fifteenth century may justly be considered as the invention of the art of printing.|ljspeech-2 wavs/LJ001-0006.mp3|And it is worth mention in passing that, as an example of fine typography,|And it is worth mention in passing that, as an example of fine typography,|ljspeech-2 wavs/LJ001-0007.mp3|the earliest book printed with movable types, the Gutenberg, or "forty-two line Bible" of about 1455,|the earliest book printed with movable types, the Gutenberg, or "forty-two line Bible" of about fourteen fifty-five,|ljspeech-3 -wavs/LJ001-0008.mp3|has never been surpassed.|has never been surpassed.|ljspeech-3 \ No newline at end of file +wavs/LJ001-0008.mp3|has never been surpassed.|has never been surpassed.|ljspeech-3 diff --git a/tests/data/ljspeech/metadata_wav.csv b/tests/data/ljspeech/metadata_wav.csv index aff73f6d40..1af6652e6a 100644 --- a/tests/data/ljspeech/metadata_wav.csv +++ b/tests/data/ljspeech/metadata_wav.csv @@ -6,4 +6,4 @@ wavs/LJ001-0004.wav|produced the block books, which were the immediate predecess wavs/LJ001-0005.wav|the invention of movable metal letters in the middle of the fifteenth century may justly be considered as the invention of the art of printing.|the invention of movable metal letters in the middle of the fifteenth century may justly be considered as the invention of the art of printing.|ljspeech-2 wavs/LJ001-0006.wav|And it is worth mention in passing that, as an example of fine typography,|And it is worth mention in passing that, as an example of fine typography,|ljspeech-2 wavs/LJ001-0007.wav|the earliest book printed with movable types, the Gutenberg, or "forty-two line Bible" of about 1455,|the earliest book printed with movable types, the Gutenberg, or "forty-two line Bible" of about fourteen fifty-five,|ljspeech-3 -wavs/LJ001-0008.wav|has never been surpassed.|has never been surpassed.|ljspeech-3 \ No newline at end of file +wavs/LJ001-0008.wav|has never been surpassed.|has never been surpassed.|ljspeech-3 diff --git a/tests/data_tests/test_loader.py b/tests/data_tests/test_loader.py index ce8738761a..252b429a16 100644 --- a/tests/data_tests/test_loader.py +++ b/tests/data_tests/test_loader.py @@ -8,7 +8,8 @@ from tests import get_tests_data_path, get_tests_output_path from TTS.tts.configs.shared_configs import BaseDatasetConfig, BaseTTSConfig -from TTS.tts.datasets import TTSDataset, load_tts_samples +from TTS.tts.datasets import load_tts_samples +from TTS.tts.datasets.dataset import TTSDataset from TTS.tts.utils.text.tokenizer import TTSTokenizer from TTS.utils.audio import AudioProcessor diff --git a/tests/inference_tests/test_synthesizer.py b/tests/inference_tests/test_synthesizer.py index ce4fc751c2..21cc194131 100644 --- a/tests/inference_tests/test_synthesizer.py +++ b/tests/inference_tests/test_synthesizer.py @@ -23,7 +23,7 @@ def test_in_out(self): tts_root_path = get_tests_input_path() tts_checkpoint = os.path.join(tts_root_path, "checkpoint_10.pth") tts_config = os.path.join(tts_root_path, "dummy_model_config.json") - synthesizer = Synthesizer(tts_checkpoint, tts_config, None, None) + synthesizer = Synthesizer(tts_checkpoint=tts_checkpoint, tts_config_path=tts_config) synthesizer.tts("Better this test works!!") def test_split_into_sentences(self): diff --git a/tests/inputs/common_voice.tsv b/tests/inputs/common_voice.tsv index 39fc4190ac..b4351d6739 100644 --- a/tests/inputs/common_voice.tsv +++ b/tests/inputs/common_voice.tsv @@ -1,6 +1,6 @@ client_id path sentence up_votes down_votes age gender accent locale segment -95324d489b122a800b840e0b0d068f7363a1a6c2cd2e7365672cc7033e38deaa794bd59edcf8196aa35c9791652b9085ac3839a98bb50ebab4a1e8538a94846b common_voice_en_20005954.mp3 The applicants are invited for coffee and visa is given immediately. 3 0 en -95324d489b122a800b840e0b0d068f7363a1a6c2cd2e7365672cc7033e38deaa794bd59edcf8196aa35c9791652b9085ac3839a98bb50ebab4a1e8538a94846b common_voice_en_20005955.mp3 Developmental robotics is related to, but differs from, evolutionary robotics. 2 0 en -95324d489b122a800b840e0b0d068f7363a1a6c2cd2e7365672cc7033e38deaa794bd59edcf8196aa35c9791652b9085ac3839a98bb50ebab4a1e8538a94846b common_voice_en_20005956.mp3 The musical was originally directed and choreographed by Alan Lund. 2 0 en -954a4181ae9fba89d1b1570f2ae148b3ee18ee2311de978e698f598db859f830d93d35574596d713518e8c96cdae01fce7a08c60c2e0a22bcf01e020924440a6 common_voice_en_19737073.mp3 He graduated from Columbia High School, in Brown County, South Dakota. 2 0 en -954a4181ae9fba89d1b1570f2ae148b3ee18ee2311de978e698f598db859f830d93d35574596d713518e8c96cdae01fce7a08c60c2e0a22bcf01e020924440a6 common_voice_en_19737074.mp3 Competition for limited resources has also resulted in some local conflicts. 2 0 en +95324d489b122a800b840e0b0d068f7363a1a6c2cd2e7365672cc7033e38deaa794bd59edcf8196aa35c9791652b9085ac3839a98bb50ebab4a1e8538a94846b common_voice_en_20005954.mp3 The applicants are invited for coffee and visa is given immediately. 3 0 en +95324d489b122a800b840e0b0d068f7363a1a6c2cd2e7365672cc7033e38deaa794bd59edcf8196aa35c9791652b9085ac3839a98bb50ebab4a1e8538a94846b common_voice_en_20005955.mp3 Developmental robotics is related to, but differs from, evolutionary robotics. 2 0 en +95324d489b122a800b840e0b0d068f7363a1a6c2cd2e7365672cc7033e38deaa794bd59edcf8196aa35c9791652b9085ac3839a98bb50ebab4a1e8538a94846b common_voice_en_20005956.mp3 The musical was originally directed and choreographed by Alan Lund. 2 0 en +954a4181ae9fba89d1b1570f2ae148b3ee18ee2311de978e698f598db859f830d93d35574596d713518e8c96cdae01fce7a08c60c2e0a22bcf01e020924440a6 common_voice_en_19737073.mp3 He graduated from Columbia High School, in Brown County, South Dakota. 2 0 en +954a4181ae9fba89d1b1570f2ae148b3ee18ee2311de978e698f598db859f830d93d35574596d713518e8c96cdae01fce7a08c60c2e0a22bcf01e020924440a6 common_voice_en_19737074.mp3 Competition for limited resources has also resulted in some local conflicts. 2 0 en diff --git a/tests/inputs/dummy_model_config.json b/tests/inputs/dummy_model_config.json index b51bb3a871..3f64c7f3df 100644 --- a/tests/inputs/dummy_model_config.json +++ b/tests/inputs/dummy_model_config.json @@ -98,5 +98,3 @@ "gst_style_tokens": 10 } } - - diff --git a/tests/inputs/language_ids.json b/tests/inputs/language_ids.json index 27bb15206f..80833d8058 100644 --- a/tests/inputs/language_ids.json +++ b/tests/inputs/language_ids.json @@ -2,4 +2,4 @@ "en": 0, "fr-fr": 1, "pt-br": 2 -} \ No newline at end of file +} diff --git a/tests/inputs/test_align_tts.json b/tests/inputs/test_align_tts.json index 3f928c7e92..80721346d5 100644 --- a/tests/inputs/test_align_tts.json +++ b/tests/inputs/test_align_tts.json @@ -155,4 +155,4 @@ "meta_file_attn_mask": null } ] -} \ No newline at end of file +} diff --git a/tests/inputs/test_speaker_encoder_config.json b/tests/inputs/test_speaker_encoder_config.json index bfcc17ab0e..ae125f1327 100644 --- a/tests/inputs/test_speaker_encoder_config.json +++ b/tests/inputs/test_speaker_encoder_config.json @@ -58,4 +58,4 @@ "storage_size": 15 // the size of the in-memory storage with respect to a single batch }, "datasets":null -} \ No newline at end of file +} diff --git a/tests/inputs/test_speedy_speech.json b/tests/inputs/test_speedy_speech.json index 4a7eea5ded..93e4790ca3 100644 --- a/tests/inputs/test_speedy_speech.json +++ b/tests/inputs/test_speedy_speech.json @@ -152,4 +152,4 @@ "meta_file_attn_mask": "tests/data/ljspeech/metadata_attn_mask.txt" } ] -} \ No newline at end of file +} diff --git a/tests/inputs/test_vocoder_audio_config.json b/tests/inputs/test_vocoder_audio_config.json index 08acc48cd3..cdf347c4eb 100644 --- a/tests/inputs/test_vocoder_audio_config.json +++ b/tests/inputs/test_vocoder_audio_config.json @@ -21,4 +21,3 @@ "do_trim_silence": false } } - diff --git a/tests/inputs/test_vocoder_multiband_melgan_config.json b/tests/inputs/test_vocoder_multiband_melgan_config.json index 82afc97727..2b6cc9e4cd 100644 --- a/tests/inputs/test_vocoder_multiband_melgan_config.json +++ b/tests/inputs/test_vocoder_multiband_melgan_config.json @@ -163,4 +163,3 @@ // PATHS "output_path": "tests/train_outputs/" } - diff --git a/tests/inputs/test_vocoder_wavegrad.json b/tests/inputs/test_vocoder_wavegrad.json index 6378c07a6d..bb06bf2448 100644 --- a/tests/inputs/test_vocoder_wavegrad.json +++ b/tests/inputs/test_vocoder_wavegrad.json @@ -113,4 +113,3 @@ // PATHS "output_path": "tests/train_outputs/" } - diff --git a/tests/inputs/test_vocoder_wavernn_config.json b/tests/inputs/test_vocoder_wavernn_config.json index ee4e5f8e42..1dd8a229f2 100644 --- a/tests/inputs/test_vocoder_wavernn_config.json +++ b/tests/inputs/test_vocoder_wavernn_config.json @@ -109,4 +109,3 @@ // PATHS "output_path": "tests/train_outputs/" } - diff --git a/tests/inputs/xtts_vocab.json b/tests/inputs/xtts_vocab.json index a3c6dcec77..e25b4e4863 100644 --- a/tests/inputs/xtts_vocab.json +++ b/tests/inputs/xtts_vocab.json @@ -12666,4 +12666,4 @@ "da kara" ] } -} \ No newline at end of file +} diff --git a/tests/text_tests/test_phonemizer.py b/tests/text_tests/test_phonemizer.py index 8810554421..f9067530e6 100644 --- a/tests/text_tests/test_phonemizer.py +++ b/tests/text_tests/test_phonemizer.py @@ -116,6 +116,12 @@ def setUp(self): output = self.phonemizer.phonemize(text, separator="") self.assertEqual(output, gt) + # UTF8 characters + text = "Åērebię" + gt = "ʑrˈɛbjɛ" + output = ESpeak("pl").phonemize(text, separator="") + self.assertEqual(output, gt) + def test_name(self): self.assertEqual(self.phonemizer.name(), "espeak") @@ -234,8 +240,12 @@ def test_is_available(self): class TestBN_Phonemizer(unittest.TestCase): def setUp(self): self.phonemizer = BN_Phonemizer() - self._TEST_CASES = "āĻ°āĻžāĻ¸ā§‚āĻ˛ā§āĻ˛ā§āĻ˛āĻžāĻš āĻ¸āĻžāĻ˛ā§āĻ˛āĻžāĻ˛ā§āĻ˛āĻžāĻšā§ āĻ†āĻ˛āĻžāĻ‡āĻšāĻŋ āĻ“ā§ŸāĻž āĻ¸āĻžāĻ˛ā§āĻ˛āĻžāĻŽ āĻļāĻŋāĻ•ā§āĻˇāĻž āĻĻāĻŋā§Ÿā§‡āĻ›ā§‡āĻ¨ āĻ¯ā§‡, āĻ•ā§‡āĻ‰ āĻ¯āĻĻāĻŋ āĻ•ā§‹āĻ¨ āĻ–āĻžāĻ°āĻžāĻĒ āĻ•āĻŋāĻ›ā§āĻ° āĻ¸āĻŽā§āĻŽā§āĻ–ā§€āĻ¨ āĻšā§Ÿ, āĻ¤āĻ–āĻ¨āĻ“ āĻ¯ā§‡āĻ¨" - self._EXPECTED = "āĻ°āĻžāĻ¸ā§‚āĻ˛ā§āĻ˛ā§āĻ˛āĻžāĻš āĻ¸āĻžāĻ˛ā§āĻ˛āĻžāĻ˛ā§āĻ˛āĻžāĻšā§ āĻ†āĻ˛āĻžāĻ‡āĻšāĻŋ āĻ“ā§ŸāĻž āĻ¸āĻžāĻ˛ā§āĻ˛āĻžāĻŽ āĻļāĻŋāĻ•ā§āĻˇāĻž āĻĻāĻŋā§Ÿā§‡āĻ›ā§‡āĻ¨ āĻ¯ā§‡ āĻ•ā§‡āĻ‰ āĻ¯āĻĻāĻŋ āĻ•ā§‹āĻ¨ āĻ–āĻžāĻ°āĻžāĻĒ āĻ•āĻŋāĻ›ā§āĻ° āĻ¸āĻŽā§āĻŽā§āĻ–ā§€āĻ¨ āĻšā§Ÿ āĻ¤āĻ–āĻ¨āĻ“ āĻ¯ā§‡āĻ¨āĨ¤" + self._TEST_CASES = ( + "āĻ°āĻžāĻ¸ā§‚āĻ˛ā§āĻ˛ā§āĻ˛āĻžāĻš āĻ¸āĻžāĻ˛ā§āĻ˛āĻžāĻ˛ā§āĻ˛āĻžāĻšā§ āĻ†āĻ˛āĻžāĻ‡āĻšāĻŋ āĻ“ā§ŸāĻž āĻ¸āĻžāĻ˛ā§āĻ˛āĻžāĻŽ āĻļāĻŋāĻ•ā§āĻˇāĻž āĻĻāĻŋā§Ÿā§‡āĻ›ā§‡āĻ¨ āĻ¯ā§‡, āĻ•ā§‡āĻ‰ āĻ¯āĻĻāĻŋ āĻ•ā§‹āĻ¨ āĻ–āĻžāĻ°āĻžāĻĒ āĻ•āĻŋāĻ›ā§āĻ° āĻ¸āĻŽā§āĻŽā§āĻ–ā§€āĻ¨ āĻšā§Ÿ, āĻ¤āĻ–āĻ¨āĻ“ āĻ¯ā§‡āĻ¨" + ) + self._EXPECTED = ( + "āĻ°āĻžāĻ¸ā§‚āĻ˛ā§āĻ˛ā§āĻ˛āĻžāĻš āĻ¸āĻžāĻ˛ā§āĻ˛āĻžāĻ˛ā§āĻ˛āĻžāĻšā§ āĻ†āĻ˛āĻžāĻ‡āĻšāĻŋ āĻ“ā§ŸāĻž āĻ¸āĻžāĻ˛ā§āĻ˛āĻžāĻŽ āĻļāĻŋāĻ•ā§āĻˇāĻž āĻĻāĻŋā§Ÿā§‡āĻ›ā§‡āĻ¨ āĻ¯ā§‡ āĻ•ā§‡āĻ‰ āĻ¯āĻĻāĻŋ āĻ•ā§‹āĻ¨ āĻ–āĻžāĻ°āĻžāĻĒ āĻ•āĻŋāĻ›ā§āĻ° āĻ¸āĻŽā§āĻŽā§āĻ–ā§€āĻ¨ āĻšā§Ÿ āĻ¤āĻ–āĻ¨āĻ“ āĻ¯ā§‡āĻ¨āĨ¤" + ) def test_phonemize(self): self.assertEqual(self.phonemizer.phonemize(self._TEST_CASES, separator=""), self._EXPECTED) diff --git a/tests/text_tests/test_text_cleaners.py b/tests/text_tests/test_text_cleaners.py index fcfa71e77d..9be1f0bf41 100644 --- a/tests/text_tests/test_text_cleaners.py +++ b/tests/text_tests/test_text_cleaners.py @@ -1,6 +1,11 @@ #!/usr/bin/env python3 -from TTS.tts.utils.text.cleaners import english_cleaners, phoneme_cleaners +from TTS.tts.utils.text.cleaners import ( + english_cleaners, + multilingual_phoneme_cleaners, + normalize_unicode, + phoneme_cleaners, +) def test_time() -> None: @@ -19,3 +24,30 @@ def test_currency() -> None: def test_expand_numbers() -> None: assert phoneme_cleaners("-1") == "minus one" assert phoneme_cleaners("1") == "one" + + +def test_multilingual_phoneme_cleaners() -> None: + assert multilingual_phoneme_cleaners("(Hello)") == "Hello" + assert multilingual_phoneme_cleaners("1:") == "1," + + +def test_normalize_unicode() -> None: + test_cases = [ + ("Häagen-Dazs", "Häagen-Dazs"), + ("äŊ åĨŊ!", "äŊ åĨŊ!"), + ("𝔄𝔅ℭâ“ĩâ“ļ⓷ī¸ˇ,ī¸¸,i⁚,i₉,㌀,Âŧ", "𝔄𝔅ℭâ“ĩâ“ļ⓷ī¸ˇ,ī¸¸,i⁚,i₉,㌀,Âŧ"), + ("Ê", "Ê"), + ("e\u0301", "Ê"), + ("a\u0300", "à"), + ("a\u0327", "aĖ§"), + ("na\u0303", "nÃŖ"), + ("o\u0302u", "ôu"), + ("n\u0303", "Ãą"), + ("\u4E2D\u56FD", "中å›Ŋ"), + ("niÃąo", "niÃąo"), + ("a\u0308", "ä"), + ("\u3053\u3093\u306b\u3061\u306f", "こんãĢãĄã¯"), + ("\u03B1\u03B2", "ιβ"), + ] + for arg, expect in test_cases: + assert normalize_unicode(arg) == expect diff --git a/tests/tts_tests/test_losses.py b/tests/tts_tests/test_losses.py index 522b7bb17c..794478dca3 100644 --- a/tests/tts_tests/test_losses.py +++ b/tests/tts_tests/test_losses.py @@ -216,7 +216,7 @@ def test_in_out(self): # pylint: disable=no-self-use late_x = -200.0 * sequence_mask(length + 1, 100).float() + 100.0 # simulate logits on late stopping loss = layer(true_x, target, length) - self.assertEqual(loss.item(), 0.0) + self.assertAlmostEqual(loss.item(), 0.0) loss = layer(early_x, target, length) self.assertAlmostEqual(loss.item(), 2.1053, places=4) diff --git a/tests/tts_tests/test_neuralhmm_tts_train.py b/tests/tts_tests/test_neuralhmm_tts_train.py index 25d9aa8148..4789d53d9e 100644 --- a/tests/tts_tests/test_neuralhmm_tts_train.py +++ b/tests/tts_tests/test_neuralhmm_tts_train.py @@ -4,7 +4,7 @@ import shutil import torch -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.neuralhmm_tts_config import NeuralhmmTTSConfig diff --git a/tests/tts_tests/test_overflow_train.py b/tests/tts_tests/test_overflow_train.py index 86fa60af72..d86bde6854 100644 --- a/tests/tts_tests/test_overflow_train.py +++ b/tests/tts_tests/test_overflow_train.py @@ -4,7 +4,7 @@ import shutil import torch -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.overflow_config import OverflowConfig diff --git a/tests/tts_tests/test_speedy_speech_train.py b/tests/tts_tests/test_speedy_speech_train.py index 530781ef88..2aac7f101d 100644 --- a/tests/tts_tests/test_speedy_speech_train.py +++ b/tests/tts_tests/test_speedy_speech_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.speedy_speech_config import SpeedySpeechConfig diff --git a/tests/tts_tests/test_tacotron2_d-vectors_train.py b/tests/tts_tests/test_tacotron2_d-vectors_train.py index 99ba4349c4..d2d1d5c35f 100644 --- a/tests/tts_tests/test_tacotron2_d-vectors_train.py +++ b/tests/tts_tests/test_tacotron2_d-vectors_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.tacotron2_config import Tacotron2Config diff --git a/tests/tts_tests/test_tacotron2_model.py b/tests/tts_tests/test_tacotron2_model.py index b1bdeb9fd1..72b6bcd46b 100644 --- a/tests/tts_tests/test_tacotron2_model.py +++ b/tests/tts_tests/test_tacotron2_model.py @@ -278,7 +278,7 @@ def test_train_step(): }, ) - batch = dict({}) + batch = {} batch["text_input"] = torch.randint(0, 24, (8, 128)).long().to(device) batch["text_lengths"] = torch.randint(100, 129, (8,)).long().to(device) batch["text_lengths"] = torch.sort(batch["text_lengths"], descending=True)[0] diff --git a/tests/tts_tests/test_tacotron2_speaker_emb_train.py b/tests/tts_tests/test_tacotron2_speaker_emb_train.py index 5f1bc3fd50..83a07d1a6c 100644 --- a/tests/tts_tests/test_tacotron2_speaker_emb_train.py +++ b/tests/tts_tests/test_tacotron2_speaker_emb_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.tacotron2_config import Tacotron2Config diff --git a/tests/tts_tests/test_tacotron2_train.py b/tests/tts_tests/test_tacotron2_train.py index 40107070e1..df0e934d8e 100644 --- a/tests/tts_tests/test_tacotron2_train.py +++ b/tests/tts_tests/test_tacotron2_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.tacotron2_config import Tacotron2Config diff --git a/tests/tts_tests/test_tacotron_model.py b/tests/tts_tests/test_tacotron_model.py index 906ec3d09f..7ec3f0df1b 100644 --- a/tests/tts_tests/test_tacotron_model.py +++ b/tests/tts_tests/test_tacotron_model.py @@ -4,6 +4,7 @@ import torch from torch import nn, optim +from trainer.generic_utils import count_parameters from tests import get_tests_input_path from TTS.tts.configs.shared_configs import CapacitronVAEConfig, GSTConfig @@ -24,11 +25,6 @@ WAV_FILE = os.path.join(get_tests_input_path(), "example_1.wav") -def count_parameters(model): - r"""Count number of trainable parameters in a network""" - return sum(p.numel() for p in model.parameters() if p.requires_grad) - - class TacotronTrainTest(unittest.TestCase): @staticmethod def test_train_step(): @@ -266,7 +262,7 @@ def test_train_step(): }, ) - batch = dict({}) + batch = {} batch["text_input"] = torch.randint(0, 24, (8, 128)).long().to(device) batch["text_lengths"] = torch.randint(100, 129, (8,)).long().to(device) batch["text_lengths"] = torch.sort(batch["text_lengths"], descending=True)[0] diff --git a/tests/tts_tests/test_tacotron_train.py b/tests/tts_tests/test_tacotron_train.py index f7751931ae..17f1fd46a6 100644 --- a/tests/tts_tests/test_tacotron_train.py +++ b/tests/tts_tests/test_tacotron_train.py @@ -2,7 +2,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.tacotron_config import TacotronConfig diff --git a/tests/tts_tests/test_vits.py b/tests/tts_tests/test_vits.py index fca9955619..c8a52e1c1b 100644 --- a/tests/tts_tests/test_vits.py +++ b/tests/tts_tests/test_vits.py @@ -13,14 +13,10 @@ Vits, VitsArgs, VitsAudioConfig, - amp_to_db, - db_to_amp, load_audio, - spec_to_mel, - wav_to_mel, - wav_to_spec, ) from TTS.tts.utils.speakers import SpeakerManager +from TTS.utils.audio.torch_transforms import amp_to_db, db_to_amp, spec_to_mel, wav_to_mel, wav_to_spec LANG_FILE = os.path.join(get_tests_input_path(), "language_ids.json") SPEAKER_ENCODER_CONFIG = os.path.join(get_tests_input_path(), "test_speaker_encoder_config.json") @@ -64,7 +60,6 @@ def test_load_audio(self): def test_dataset(self): """TODO:""" - ... def test_init_multispeaker(self): num_speakers = 10 @@ -213,7 +208,7 @@ def test_d_vector_forward(self): d_vector_file=[os.path.join(get_tests_data_path(), "dummy_speakers.json")], ) config = VitsConfig(model_args=args) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) model.train() input_dummy, input_lengths, _, spec, spec_lengths, waveform = self._create_inputs(config, batch_size=batch_size) d_vectors = torch.randn(batch_size, 256).to(device) @@ -358,7 +353,7 @@ def test_d_vector_inference(self): d_vector_file=[os.path.join(get_tests_data_path(), "dummy_speakers.json")], ) config = VitsConfig(model_args=args) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) model.eval() # batch size = 1 input_dummy = torch.randint(0, 24, (1, 128)).long().to(device) @@ -512,7 +507,7 @@ def test_train_step_upsampling_interpolation(self): def test_train_eval_log(self): batch_size = 2 config = VitsConfig(model_args=VitsArgs(num_chars=32, spec_segment_size=10)) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) model.run_data_dep_init = False model.train() batch = self._create_batch(config, batch_size) @@ -531,7 +526,7 @@ def test_train_eval_log(self): def test_test_run(self): config = VitsConfig(model_args=VitsArgs(num_chars=32)) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) model.run_data_dep_init = False model.eval() test_figures, test_audios = model.test_run(None) @@ -541,7 +536,7 @@ def test_test_run(self): def test_load_checkpoint(self): chkp_path = os.path.join(get_tests_output_path(), "dummy_glow_tts_checkpoint.pth") config = VitsConfig(VitsArgs(num_chars=32)) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) chkp = {} chkp["model"] = model.state_dict() torch.save(chkp, chkp_path) @@ -552,20 +547,20 @@ def test_load_checkpoint(self): def test_get_criterion(self): config = VitsConfig(VitsArgs(num_chars=32)) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) criterion = model.get_criterion() self.assertTrue(criterion is not None) def test_init_from_config(self): config = VitsConfig(model_args=VitsArgs(num_chars=32)) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) config = VitsConfig(model_args=VitsArgs(num_chars=32, num_speakers=2)) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) self.assertTrue(not hasattr(model, "emb_g")) config = VitsConfig(model_args=VitsArgs(num_chars=32, num_speakers=2, use_speaker_embedding=True)) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) self.assertEqual(model.num_speakers, 2) self.assertTrue(hasattr(model, "emb_g")) @@ -577,7 +572,7 @@ def test_init_from_config(self): speakers_file=os.path.join(get_tests_data_path(), "ljspeech", "speakers.json"), ) ) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) self.assertEqual(model.num_speakers, 10) self.assertTrue(hasattr(model, "emb_g")) @@ -589,7 +584,7 @@ def test_init_from_config(self): d_vector_file=[os.path.join(get_tests_data_path(), "dummy_speakers.json")], ) ) - model = Vits.init_from_config(config, verbose=False).to(device) + model = Vits.init_from_config(config).to(device) self.assertTrue(model.num_speakers == 1) self.assertTrue(not hasattr(model, "emb_g")) self.assertTrue(model.embedded_speaker_dim == config.d_vector_dim) diff --git a/tests/tts_tests/test_vits_multilingual_speaker_emb_train.py b/tests/tts_tests/test_vits_multilingual_speaker_emb_train.py index 71597ef32f..09df7d29f2 100644 --- a/tests/tts_tests/test_vits_multilingual_speaker_emb_train.py +++ b/tests/tts_tests/test_vits_multilingual_speaker_emb_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.config.shared_configs import BaseDatasetConfig diff --git a/tests/tts_tests/test_vits_multilingual_train-d_vectors.py b/tests/tts_tests/test_vits_multilingual_train-d_vectors.py index fd58db534a..7ae09c0e5c 100644 --- a/tests/tts_tests/test_vits_multilingual_train-d_vectors.py +++ b/tests/tts_tests/test_vits_multilingual_train-d_vectors.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.config.shared_configs import BaseDatasetConfig diff --git a/tests/tts_tests/test_vits_speaker_emb_train.py b/tests/tts_tests/test_vits_speaker_emb_train.py index b7fe197cfe..69fae21f8d 100644 --- a/tests/tts_tests/test_vits_speaker_emb_train.py +++ b/tests/tts_tests/test_vits_speaker_emb_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.vits_config import VitsConfig diff --git a/tests/tts_tests/test_vits_train.py b/tests/tts_tests/test_vits_train.py index ea5dc02405..78f42d154b 100644 --- a/tests/tts_tests/test_vits_train.py +++ b/tests/tts_tests/test_vits_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.vits_config import VitsConfig diff --git a/tests/tts_tests2/test_align_tts_train.py b/tests/tts_tests2/test_align_tts_train.py index 9b0b730df4..91c3c35bc6 100644 --- a/tests/tts_tests2/test_align_tts_train.py +++ b/tests/tts_tests2/test_align_tts_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.align_tts_config import AlignTTSConfig diff --git a/tests/tts_tests2/test_delightful_tts_d-vectors_train.py b/tests/tts_tests2/test_delightful_tts_d-vectors_train.py index 8fc4ea7e9b..1e5cd49f73 100644 --- a/tests/tts_tests2/test_delightful_tts_d-vectors_train.py +++ b/tests/tts_tests2/test_delightful_tts_d-vectors_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.delightful_tts_config import DelightfulTtsAudioConfig, DelightfulTTSConfig diff --git a/tests/tts_tests2/test_delightful_tts_emb_spk.py b/tests/tts_tests2/test_delightful_tts_emb_spk.py index 6fb70c5f61..9bbf7a55ea 100644 --- a/tests/tts_tests2/test_delightful_tts_emb_spk.py +++ b/tests/tts_tests2/test_delightful_tts_emb_spk.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.delightful_tts_config import DelightfulTtsAudioConfig, DelightfulTTSConfig diff --git a/tests/tts_tests2/test_delightful_tts_train.py b/tests/tts_tests2/test_delightful_tts_train.py index a917d77657..3e6fbd2e86 100644 --- a/tests/tts_tests2/test_delightful_tts_train.py +++ b/tests/tts_tests2/test_delightful_tts_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.config.shared_configs import BaseAudioConfig diff --git a/tests/tts_tests2/test_fast_pitch_speaker_emb_train.py b/tests/tts_tests2/test_fast_pitch_speaker_emb_train.py index 7f79bfcab2..e6bc9f9feb 100644 --- a/tests/tts_tests2/test_fast_pitch_speaker_emb_train.py +++ b/tests/tts_tests2/test_fast_pitch_speaker_emb_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.config.shared_configs import BaseAudioConfig diff --git a/tests/tts_tests2/test_fast_pitch_train.py b/tests/tts_tests2/test_fast_pitch_train.py index a525715b53..fe87c8b600 100644 --- a/tests/tts_tests2/test_fast_pitch_train.py +++ b/tests/tts_tests2/test_fast_pitch_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.config.shared_configs import BaseAudioConfig diff --git a/tests/tts_tests2/test_fastspeech_2_speaker_emb_train.py b/tests/tts_tests2/test_fastspeech_2_speaker_emb_train.py index 35bda597d5..735d2fc4c6 100644 --- a/tests/tts_tests2/test_fastspeech_2_speaker_emb_train.py +++ b/tests/tts_tests2/test_fastspeech_2_speaker_emb_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.config.shared_configs import BaseAudioConfig diff --git a/tests/tts_tests2/test_fastspeech_2_train.py b/tests/tts_tests2/test_fastspeech_2_train.py index dd4b07d240..07fc5a1a2c 100644 --- a/tests/tts_tests2/test_fastspeech_2_train.py +++ b/tests/tts_tests2/test_fastspeech_2_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.config.shared_configs import BaseAudioConfig diff --git a/tests/tts_tests2/test_forward_tts.py b/tests/tts_tests2/test_forward_tts.py index cec0f211c8..13a2c270af 100644 --- a/tests/tts_tests2/test_forward_tts.py +++ b/tests/tts_tests2/test_forward_tts.py @@ -6,29 +6,7 @@ # pylint: disable=unused-variable -def expand_encoder_outputs_test(): - model = ForwardTTS(ForwardTTSArgs(num_chars=10)) - - inputs = T.rand(2, 5, 57) - durations = T.randint(1, 4, (2, 57)) - - x_mask = T.ones(2, 1, 57) - y_mask = T.ones(2, 1, durations.sum(1).max()) - - expanded, _ = model.expand_encoder_outputs(inputs, durations, x_mask, y_mask) - - for b in range(durations.shape[0]): - index = 0 - for idx, dur in enumerate(durations[b]): - diff = ( - expanded[b, :, index : index + dur.item()] - - inputs[b, :, idx].repeat(dur.item()).view(expanded[b, :, index : index + dur.item()].shape) - ).sum() - assert abs(diff) < 1e-6, diff - index += dur - - -def model_input_output_test(): +def test_model_input_output(): """Assert the output shapes of the model in different modes""" # VANILLA MODEL diff --git a/tests/tts_tests2/test_glow_tts.py b/tests/tts_tests2/test_glow_tts.py index 2a723f105f..3c7ac51556 100644 --- a/tests/tts_tests2/test_glow_tts.py +++ b/tests/tts_tests2/test_glow_tts.py @@ -4,6 +4,7 @@ import torch from torch import optim +from trainer.generic_utils import count_parameters from trainer.logging.tensorboard_logger import TensorboardLogger from tests import get_tests_data_path, get_tests_input_path, get_tests_output_path @@ -26,11 +27,6 @@ BATCH_SIZE = 3 -def count_parameters(model): - r"""Count number of trainable parameters in a network""" - return sum(p.numel() for p in model.parameters() if p.requires_grad) - - class TestGlowTTS(unittest.TestCase): @staticmethod def _create_inputs(batch_size=8): @@ -136,7 +132,7 @@ def _test_forward_with_d_vector(self, batch_size): d_vector_dim=256, d_vector_file=os.path.join(get_tests_data_path(), "dummy_speakers.json"), ) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) model.train() print(" > Num parameters for GlowTTS model:%s" % (count_parameters(model))) # inference encoder and decoder with MAS @@ -162,7 +158,7 @@ def _test_forward_with_speaker_id(self, batch_size): use_speaker_embedding=True, num_speakers=24, ) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) model.train() print(" > Num parameters for GlowTTS model:%s" % (count_parameters(model))) # inference encoder and decoder with MAS @@ -210,7 +206,7 @@ def _test_inference_with_d_vector(self, batch_size): d_vector_dim=256, d_vector_file=os.path.join(get_tests_data_path(), "dummy_speakers.json"), ) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) model.eval() outputs = model.inference(input_dummy, {"x_lengths": input_lengths, "d_vectors": d_vector}) self._assert_inference_outputs(outputs, input_dummy, mel_spec) @@ -228,7 +224,7 @@ def _test_inference_with_speaker_ids(self, batch_size): use_speaker_embedding=True, num_speakers=24, ) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) outputs = model.inference(input_dummy, {"x_lengths": input_lengths, "speaker_ids": speaker_ids}) self._assert_inference_outputs(outputs, input_dummy, mel_spec) @@ -303,7 +299,7 @@ def test_train_eval_log(self): batch["d_vectors"] = None batch["speaker_ids"] = None config = GlowTTSConfig(num_chars=32) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) model.run_data_dep_init = False model.train() logger = TensorboardLogger( @@ -317,7 +313,7 @@ def test_train_eval_log(self): def test_test_run(self): config = GlowTTSConfig(num_chars=32) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) model.run_data_dep_init = False model.eval() test_figures, test_audios = model.test_run(None) @@ -327,7 +323,7 @@ def test_test_run(self): def test_load_checkpoint(self): chkp_path = os.path.join(get_tests_output_path(), "dummy_glow_tts_checkpoint.pth") config = GlowTTSConfig(num_chars=32) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) chkp = {} chkp["model"] = model.state_dict() torch.save(chkp, chkp_path) @@ -338,21 +334,21 @@ def test_load_checkpoint(self): def test_get_criterion(self): config = GlowTTSConfig(num_chars=32) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) criterion = model.get_criterion() self.assertTrue(criterion is not None) def test_init_from_config(self): config = GlowTTSConfig(num_chars=32) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) config = GlowTTSConfig(num_chars=32, num_speakers=2) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) self.assertTrue(model.num_speakers == 2) self.assertTrue(not hasattr(model, "emb_g")) config = GlowTTSConfig(num_chars=32, num_speakers=2, use_speaker_embedding=True) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) self.assertTrue(model.num_speakers == 2) self.assertTrue(hasattr(model, "emb_g")) @@ -362,7 +358,7 @@ def test_init_from_config(self): use_speaker_embedding=True, speakers_file=os.path.join(get_tests_data_path(), "ljspeech", "speakers.json"), ) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) self.assertTrue(model.num_speakers == 10) self.assertTrue(hasattr(model, "emb_g")) @@ -372,7 +368,7 @@ def test_init_from_config(self): d_vector_dim=256, d_vector_file=os.path.join(get_tests_data_path(), "dummy_speakers.json"), ) - model = GlowTTS.init_from_config(config, verbose=False).to(device) + model = GlowTTS.init_from_config(config).to(device) self.assertTrue(model.num_speakers == 1) self.assertTrue(not hasattr(model, "emb_g")) self.assertTrue(model.c_in_channels == config.d_vector_dim) diff --git a/tests/tts_tests2/test_glow_tts_d-vectors_train.py b/tests/tts_tests2/test_glow_tts_d-vectors_train.py index f1cfd4368f..8236607c25 100644 --- a/tests/tts_tests2/test_glow_tts_d-vectors_train.py +++ b/tests/tts_tests2/test_glow_tts_d-vectors_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.glow_tts_config import GlowTTSConfig diff --git a/tests/tts_tests2/test_glow_tts_speaker_emb_train.py b/tests/tts_tests2/test_glow_tts_speaker_emb_train.py index b1eb6237a4..4a8bd0658d 100644 --- a/tests/tts_tests2/test_glow_tts_speaker_emb_train.py +++ b/tests/tts_tests2/test_glow_tts_speaker_emb_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.glow_tts_config import GlowTTSConfig diff --git a/tests/tts_tests2/test_glow_tts_train.py b/tests/tts_tests2/test_glow_tts_train.py index 0a8e226b65..1d7f913575 100644 --- a/tests/tts_tests2/test_glow_tts_train.py +++ b/tests/tts_tests2/test_glow_tts_train.py @@ -3,7 +3,7 @@ import os import shutil -from trainer import get_last_checkpoint +from trainer.io import get_last_checkpoint from tests import get_device_id, get_tests_output_path, run_cli from TTS.tts.configs.glow_tts_config import GlowTTSConfig diff --git a/tests/vc_tests/test_freevc.py b/tests/vc_tests/test_freevc.py index a4a4f72679..fe07b2723c 100644 --- a/tests/vc_tests/test_freevc.py +++ b/tests/vc_tests/test_freevc.py @@ -2,10 +2,10 @@ import unittest import torch +from trainer.generic_utils import count_parameters from tests import get_tests_input_path -from TTS.vc.configs.freevc_config import FreeVCConfig -from TTS.vc.models.freevc import FreeVC +from TTS.vc.models.freevc import FreeVC, FreeVCConfig # pylint: disable=unused-variable # pylint: disable=no-self-use @@ -20,38 +20,21 @@ BATCH_SIZE = 3 -def count_parameters(model): - r"""Count number of trainable parameters in a network""" - return sum(p.numel() for p in model.parameters() if p.requires_grad) - - class TestFreeVC(unittest.TestCase): def _create_inputs(self, config, batch_size=2): - input_dummy = torch.rand(batch_size, 30 * config.audio["hop_length"]).to(device) - input_lengths = torch.randint(100, 30 * config.audio["hop_length"], (batch_size,)).long().to(device) - input_lengths[-1] = 30 * config.audio["hop_length"] spec = torch.rand(batch_size, 30, config.audio["filter_length"] // 2 + 1).to(device) mel = torch.rand(batch_size, 30, config.audio["n_mel_channels"]).to(device) spec_lengths = torch.randint(20, 30, (batch_size,)).long().to(device) spec_lengths[-1] = spec.size(2) waveform = torch.rand(batch_size, spec.size(2) * config.audio["hop_length"]).to(device) - return input_dummy, input_lengths, mel, spec, spec_lengths, waveform + return mel, spec, spec_lengths, waveform @staticmethod def _create_inputs_inference(): - source_wav = torch.rand(16000) + source_wav = torch.rand(15999) target_wav = torch.rand(16000) return source_wav, target_wav - @staticmethod - def _check_parameter_changes(model, model_ref): - count = 0 - for param, param_ref in zip(model.parameters(), model_ref.parameters()): - assert (param != param_ref).any(), "param {} with shape {} not updated!! \n{}\n{}".format( - count, param.shape, param, param_ref - ) - count += 1 - def test_methods(self): config = FreeVCConfig() model = FreeVC(config).to(device) @@ -74,7 +57,7 @@ def _test_forward(self, batch_size): model.train() print(" > Num parameters for FreeVC model:%s" % (count_parameters(model))) - _, _, mel, spec, spec_lengths, waveform = self._create_inputs(config, batch_size) + mel, spec, spec_lengths, waveform = self._create_inputs(config, batch_size) wavlm_vec = model.extract_wavlm_features(waveform) wavlm_vec_lengths = torch.ones(batch_size, dtype=torch.long) @@ -91,7 +74,7 @@ def _test_inference(self, batch_size): model = FreeVC(config).to(device) model.eval() - _, _, mel, _, _, waveform = self._create_inputs(config, batch_size) + mel, _, _, waveform = self._create_inputs(config, batch_size) wavlm_vec = model.extract_wavlm_features(waveform) wavlm_vec_lengths = torch.ones(batch_size, dtype=torch.long) @@ -113,23 +96,17 @@ def test_voice_conversion(self): source_wav, target_wav = self._create_inputs_inference() output_wav = model.voice_conversion(source_wav, target_wav) assert ( - output_wav.shape[0] + config.audio.hop_length == source_wav.shape[0] - ), f"{output_wav.shape} != {source_wav.shape}" + output_wav.shape[0] == source_wav.shape[0] - source_wav.shape[0] % config.audio.hop_length + ), f"{output_wav.shape} != {source_wav.shape}, {config.audio.hop_length}" - def test_train_step(self): - ... + def test_train_step(self): ... - def test_train_eval_log(self): - ... + def test_train_eval_log(self): ... - def test_test_run(self): - ... + def test_test_run(self): ... - def test_load_checkpoint(self): - ... + def test_load_checkpoint(self): ... - def test_get_criterion(self): - ... + def test_get_criterion(self): ... - def test_init_from_config(self): - ... + def test_init_from_config(self): ... diff --git a/tests/vc_tests/test_openvoice.py b/tests/vc_tests/test_openvoice.py new file mode 100644 index 0000000000..c9f7ae3931 --- /dev/null +++ b/tests/vc_tests/test_openvoice.py @@ -0,0 +1,42 @@ +import os +import unittest + +import torch + +from tests import get_tests_input_path +from TTS.vc.models.openvoice import OpenVoice, OpenVoiceConfig + +torch.manual_seed(1) +use_cuda = torch.cuda.is_available() +device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") + +c = OpenVoiceConfig() + +WAV_FILE = os.path.join(get_tests_input_path(), "example_1.wav") + + +class TestOpenVoice(unittest.TestCase): + + @staticmethod + def _create_inputs_inference(): + source_wav = torch.rand(16100) + target_wav = torch.rand(16000) + return source_wav, target_wav + + def test_load_audio(self): + config = OpenVoiceConfig() + model = OpenVoice(config).to(device) + wav = model.load_audio(WAV_FILE) + wav2 = model.load_audio(wav) + assert all(torch.isclose(wav, wav2)) + + def test_voice_conversion(self): + config = OpenVoiceConfig() + model = OpenVoice(config).to(device) + model.eval() + + source_wav, target_wav = self._create_inputs_inference() + output_wav = model.voice_conversion(source_wav, target_wav) + assert ( + output_wav.shape[0] == source_wav.shape[0] - source_wav.shape[0] % config.audio.hop_length + ), f"{output_wav.shape} != {source_wav.shape}" diff --git a/tests/vocoder_tests/test_wavegrad_train.py b/tests/vocoder_tests/test_wavegrad_train.py index fe56ee783f..9b10759505 100644 --- a/tests/vocoder_tests/test_wavegrad_train.py +++ b/tests/vocoder_tests/test_wavegrad_train.py @@ -1,43 +1,54 @@ import glob import os import shutil +import unittest from tests import get_device_id, get_tests_output_path, run_cli from TTS.vocoder.configs import WavegradConfig -config_path = os.path.join(get_tests_output_path(), "test_vocoder_config.json") -output_path = os.path.join(get_tests_output_path(), "train_outputs") - -config = WavegradConfig( - batch_size=8, - eval_batch_size=8, - num_loader_workers=0, - num_eval_loader_workers=0, - run_eval=True, - test_delay_epochs=-1, - epochs=1, - seq_len=8192, - eval_split_size=1, - print_step=1, - print_eval=True, - data_path="tests/data/ljspeech", - output_path=output_path, - test_noise_schedule={"min_val": 1e-6, "max_val": 1e-2, "num_steps": 2}, -) -config.audio.do_trim_silence = True -config.audio.trim_db = 60 -config.save_json(config_path) - -# train the model for one epoch -command_train = f"CUDA_VISIBLE_DEVICES='{get_device_id()}' python TTS/bin/train_vocoder.py --config_path {config_path} " -run_cli(command_train) - -# Find latest folder -continue_path = max(glob.glob(os.path.join(output_path, "*/")), key=os.path.getmtime) - -# restore the model and continue training for one more epoch -command_train = ( - f"CUDA_VISIBLE_DEVICES='{get_device_id()}' python TTS/bin/train_vocoder.py --continue_path {continue_path} " -) -run_cli(command_train) -shutil.rmtree(continue_path) + +class WavegradTrainingTest(unittest.TestCase): + # TODO: Reactivate after improving CI run times + # This test currently takes ~2h on CI (15min/step vs 8sec/step locally) + if os.getenv("GITHUB_ACTIONS") == "true": + __test__ = False + + def test_train(self): # pylint: disable=no-self-use + config_path = os.path.join(get_tests_output_path(), "test_vocoder_config.json") + output_path = os.path.join(get_tests_output_path(), "train_outputs") + + config = WavegradConfig( + batch_size=8, + eval_batch_size=8, + num_loader_workers=0, + num_eval_loader_workers=0, + run_eval=True, + test_delay_epochs=-1, + epochs=1, + seq_len=8192, + eval_split_size=1, + print_step=1, + print_eval=True, + data_path="tests/data/ljspeech", + output_path=output_path, + test_noise_schedule={"min_val": 1e-6, "max_val": 1e-2, "num_steps": 2}, + ) + config.audio.do_trim_silence = True + config.audio.trim_db = 60 + config.save_json(config_path) + + # train the model for one epoch + command_train = ( + f"CUDA_VISIBLE_DEVICES='{get_device_id()}' python TTS/bin/train_vocoder.py --config_path {config_path} " + ) + run_cli(command_train) + + # Find latest folder + continue_path = max(glob.glob(os.path.join(output_path, "*/")), key=os.path.getmtime) + + # restore the model and continue training for one more epoch + command_train = ( + f"CUDA_VISIBLE_DEVICES='{get_device_id()}' python TTS/bin/train_vocoder.py --continue_path {continue_path} " + ) + run_cli(command_train) + shutil.rmtree(continue_path) diff --git a/tests/xtts_tests/test_xtts_gpt_train.py b/tests/xtts_tests/test_xtts_gpt_train.py index b8b9a4e388..bb592f1f2d 100644 --- a/tests/xtts_tests/test_xtts_gpt_train.py +++ b/tests/xtts_tests/test_xtts_gpt_train.py @@ -8,7 +8,8 @@ from TTS.config.shared_configs import BaseDatasetConfig from TTS.tts.datasets import load_tts_samples from TTS.tts.layers.xtts.dvae import DiscreteVAE -from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig, XttsAudioConfig +from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig +from TTS.tts.models.xtts import XttsAudioConfig config_dataset = BaseDatasetConfig( formatter="ljspeech", diff --git a/tests/xtts_tests/test_xtts_v2-0_gpt_train.py b/tests/xtts_tests/test_xtts_v2-0_gpt_train.py index 6663433c12..454e867385 100644 --- a/tests/xtts_tests/test_xtts_v2-0_gpt_train.py +++ b/tests/xtts_tests/test_xtts_v2-0_gpt_train.py @@ -8,7 +8,8 @@ from TTS.config.shared_configs import BaseDatasetConfig from TTS.tts.datasets import load_tts_samples from TTS.tts.layers.xtts.dvae import DiscreteVAE -from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig, XttsAudioConfig +from TTS.tts.layers.xtts.trainer.gpt_trainer import GPTArgs, GPTTrainer, GPTTrainerConfig +from TTS.tts.models.xtts import XttsAudioConfig config_dataset = BaseDatasetConfig( formatter="ljspeech", diff --git a/tests/zoo_tests/test_models.py b/tests/zoo_tests/test_models.py index 8fa56e287a..461b4fbe12 100644 --- a/tests/zoo_tests/test_models.py +++ b/tests/zoo_tests/test_models.py @@ -1,14 +1,13 @@ #!/usr/bin/env python3` -import glob import os import shutil import torch +from trainer.io import get_user_data_dir from tests import get_tests_data_path, get_tests_output_path, run_cli from TTS.tts.utils.languages import LanguageManager from TTS.tts.utils.speakers import SpeakerManager -from TTS.utils.generic_utils import get_user_data_dir from TTS.utils.manage import ModelManager MODELS_WITH_SEP_TESTS = [ @@ -30,34 +29,31 @@ def run_models(offset=0, step=1): print(f"\n > Run - {model_name}") model_path, _, _ = manager.download_model(model_name) if "tts_models" in model_name: - local_download_dir = os.path.dirname(model_path) + local_download_dir = model_path.parent # download and run the model - speaker_files = glob.glob(local_download_dir + "/speaker*") - language_files = glob.glob(local_download_dir + "/language*") - language_id = "" + speaker_files = list(local_download_dir.glob("speaker*")) + language_files = list(local_download_dir.glob("language*")) + speaker_arg = "" + language_arg = "" if len(speaker_files) > 0: # multi-speaker model - if "speaker_ids" in speaker_files[0]: + if "speaker_ids" in speaker_files[0].stem: speaker_manager = SpeakerManager(speaker_id_file_path=speaker_files[0]) - elif "speakers" in speaker_files[0]: + elif "speakers" in speaker_files[0].stem: speaker_manager = SpeakerManager(d_vectors_file_path=speaker_files[0]) - - # multi-lingual model - Assuming multi-lingual models are also multi-speaker - if len(language_files) > 0 and "language_ids" in language_files[0]: - language_manager = LanguageManager(language_ids_file_path=language_files[0]) - language_id = language_manager.language_names[0] - - speaker_id = list(speaker_manager.name_to_id.keys())[0] - run_cli( - f"tts --model_name {model_name} " - f'--text "This is an example." --out_path "{output_path}" --speaker_idx "{speaker_id}" --language_idx "{language_id}" --progress_bar False' - ) - else: - # single-speaker model - run_cli( - f"tts --model_name {model_name} " - f'--text "This is an example." --out_path "{output_path}" --progress_bar False' - ) + speakers = list(speaker_manager.name_to_id.keys()) + if len(speakers) > 1: + speaker_arg = f'--speaker_idx "{speakers[0]}"' + if len(language_files) > 0 and "language_ids" in language_files[0].stem: + # multi-lingual model + language_manager = LanguageManager(language_ids_file_path=language_files[0]) + languages = language_manager.language_names + if len(languages) > 1: + language_arg = f'--language_idx "{languages[0]}"' + run_cli( + f'tts --model_name {model_name} --text "This is an example." ' + f'--out_path "{output_path}" {speaker_arg} {language_arg} --no-progress_bar' + ) # remove downloaded models shutil.rmtree(local_download_dir) shutil.rmtree(get_user_data_dir("tts")) @@ -66,7 +62,7 @@ def run_models(offset=0, step=1): reference_wav = os.path.join(get_tests_data_path(), "ljspeech", "wavs", "LJ001-0032.wav") run_cli( f"tts --model_name {model_name} " - f'--out_path "{output_path}" --source_wav "{speaker_wav}" --target_wav "{reference_wav}" --progress_bar False' + f'--out_path "{output_path}" --source_wav "{speaker_wav}" --target_wav "{reference_wav}" --no-progress_bar' ) else: # only download the model @@ -83,14 +79,14 @@ def test_xtts(): run_cli( "yes | " f"tts --model_name tts_models/multilingual/multi-dataset/xtts_v1.1 " - f'--text "This is an example." --out_path "{output_path}" --progress_bar False --use_cuda True ' + f'--text "This is an example." --out_path "{output_path}" --no-progress_bar --use_cuda ' f'--speaker_wav "{speaker_wav}" --language_idx "en"' ) else: run_cli( "yes | " f"tts --model_name tts_models/multilingual/multi-dataset/xtts_v1.1 " - f'--text "This is an example." --out_path "{output_path}" --progress_bar False ' + f'--text "This is an example." --out_path "{output_path}" --no-progress_bar ' f'--speaker_wav "{speaker_wav}" --language_idx "en"' ) @@ -138,14 +134,14 @@ def test_xtts_v2(): run_cli( "yes | " f"tts --model_name tts_models/multilingual/multi-dataset/xtts_v2 " - f'--text "This is an example." --out_path "{output_path}" --progress_bar False --use_cuda True ' + f'--text "This is an example." --out_path "{output_path}" --no-progress_bar --use_cuda ' f'--speaker_wav "{speaker_wav}" "{speaker_wav_2}" --language_idx "en"' ) else: run_cli( "yes | " f"tts --model_name tts_models/multilingual/multi-dataset/xtts_v2 " - f'--text "This is an example." --out_path "{output_path}" --progress_bar False ' + f'--text "This is an example." --out_path "{output_path}" --no-progress_bar ' f'--speaker_wav "{speaker_wav}" "{speaker_wav_2}" --language_idx "en"' ) @@ -215,12 +211,12 @@ def test_tortoise(): if use_gpu: run_cli( f" tts --model_name tts_models/en/multi-dataset/tortoise-v2 " - f'--text "This is an example." --out_path "{output_path}" --progress_bar False --use_cuda True' + f'--text "This is an example." --out_path "{output_path}" --no-progress_bar --use_cuda' ) else: run_cli( f" tts --model_name tts_models/en/multi-dataset/tortoise-v2 " - f'--text "This is an example." --out_path "{output_path}" --progress_bar False' + f'--text "This is an example." --out_path "{output_path}" --no-progress_bar' ) @@ -231,12 +227,12 @@ def test_bark(): if use_gpu: run_cli( f" tts --model_name tts_models/multilingual/multi-dataset/bark " - f'--text "This is an example." --out_path "{output_path}" --progress_bar False --use_cuda True' + f'--text "This is an example." --out_path "{output_path}" --no-progress_bar --use_cuda' ) else: run_cli( f" tts --model_name tts_models/multilingual/multi-dataset/bark " - f'--text "This is an example." --out_path "{output_path}" --progress_bar False' + f'--text "This is an example." --out_path "{output_path}" --no-progress_bar' ) @@ -249,7 +245,7 @@ def test_voice_conversion(): output_path = os.path.join(get_tests_output_path(), "output.wav") run_cli( f"tts --model_name {model_name}" - f" --out_path {output_path} --speaker_wav {speaker_wav} --reference_wav {reference_wav} --language_idx {language_id} --progress_bar False" + f" --out_path {output_path} --speaker_wav {speaker_wav} --reference_wav {reference_wav} --language_idx {language_id} --no-progress_bar" )