ASRTK: Automatic Speech Recognition Toolkit

ASRTK is a Python toolkit focused on collecting and processing high-quality audio datasets for training Automatic Speech Recognition (ASR) systems. It provides a streamlined workflow and utilities for downloading, cleaning, and preparing audio data with accurate transcriptions.

Key Features

Data Collection & Preparation

• YouTube Integration: Automated download of videos with subtitles from playlists
• Audio Processing: Advanced audio splitting, conversion, and resampling capabilities
• Subtitle Processing: Comprehensive VTT file handling with support for cleaning and normalization

Text Processing & Analysis

• Text Normalization: Intelligent handling of abbreviations, numbers, and special characters
• Pattern Analysis: Tools for analyzing word frequencies and text patterns

Audio Processing

• Forced Alignment: Precise audio-text alignment using Silero VAD models
• Format Optimization: Efficient audio format conversion with multi-threading support
• Quality Control: Advanced handling of silence detection and failed alignments

Workflow Automation

• Batch Processing: Efficient handling of large datasets
• Pipeline Integration: Seamless workflow from data collection to model training
• Progress Tracking: Detailed logging and progress monitoring

Requirements

• Python 3.9 or higher
• PyTorch 1.7 or higher
• NumPy
• SciPy
• librosa
• pydub
• transformers
• BeautifulSoup4
• Additional dependencies in requirements.txt

Installation

# Clone the repository
git clone https://github.com/ysdede/asrtk.git

# Navigate to project directory
cd asrtk

# Install dependencies
pip install -r requirements.txt

Quick Start Guide

1. Download Training Data

# Download from a single playlist
asrtk download-playlist ./dataset "https://www.youtube.com/playlist?list=PLAYLIST_ID"

# Or download from multiple playlists
asrtk download-playlist ./dataset --playlist-file playlists.txt

2. Process Audio Files

# Split audio into chunks based on subtitles
asrtk split ./dataset ./chunks --tolerance 500 -fm --restore-punctuation

# Convert to optimized format
asrtk convert-opus ./chunks --remove-original --workers 8

3. Clean and Normalize Text

# Apply text normalization rules
asrtk apply-patch rules.json ./chunks

# Analyze text patterns
asrtk create-wordset ./chunks --min-frequency 5

Advanced Usage

Audio Processing Options

asrtk split INPUT_DIR OUTPUT_DIR [OPTIONS]
  --tolerance MILLISECONDS    Alignment tolerance
  --force-merge, -fm         Merge consecutive subtitles
  --restore-punctuation      Use BERT for punctuation restoration
  --keep-effects            Preserve effect annotations

Text Processing Features

asrtk apply-patch RULES INPUT_DIR [OPTIONS]
  --dry-run                 Preview changes
  --show-replacements       Display detailed replacement plan
  --verbose                Show per-file changes

Batch Processing

asrtk convert-opus INPUT_DIR [OPTIONS]
  --input-type FORMAT       Source audio format
  --remove-original        Delete source files after conversion
  --workers NUMBER         Parallel processing threads

Project Structure

asrtk/
├── dataset/               # Downloaded content
│   ├── json_info/        # Metadata cache
│   └── content/          # Audio and subtitle files
├── chunks/               # Processed audio segments
└── output/              # Final processed dataset

Contributing

We welcome contributions! Please feel free to submit pull requests or open issues for:

• Bug fixes
• Feature additions
• Documentation improvements
• Performance optimizations

License

This project is licensed under the MIT License - see the LICENSE file for details.