Generating Music with Deep Learning

This repository contains all the scripts used in the project Generating Music with Deep Learning. It is divided into two folders:

• midinet. It covers the second part of the project.
• preliminary_work. It includes the preprocessing steps for MAESTRO dataset, as well as the codes for training the LSTM, GAN and VAE models.

In every folder you will find additional information.

---

Spoken and Written Language Processing, GCiED

Alex Carrillo, Pau Magariño, Guillermo Mora and Robert Tura

-June 19, 2020, Universitat Politècnica de Catalunya

Abstract. This paper presents various approaches of music generation based on deep learning techniques. First, after a short introduction to the topic, the article analyses three early proposals: a LSTM, a Variational Autoencoder and a GAN, for symbolic music generation on MIDI, and we gently illustrate the variety of concerns they entail. Second, we leverage the power of a novel architecture called MidiNet by conducting our own refinements, and give examples of how this model, equipped with such a conditional mechanism on melody, can achieve very promising results on music generation.

Contents

1. Aim and motivation
2. Introduction
   1. MIDI
   2. Piano roll
   3. Dataset and Preprocessing
3. First approach
   1. First models
      1. LSTM
      2. Variational Autoencoder
      3. GAN
   2. Main findings
4. MidiNet
   1. Overview
   2. Dataset and Preprocessing
   3. Architecture details
   4. Errors corrected
5. Proposed models
   1. MidiNet Baseline
   2. Embeddings
   3. Spectral normalization
   4. Previous data to the discriminator
   5. MLP discriminator
   6. Embeddings with previous data plus spectral normalization
6. Results comparison
   1. Baseline
   2. Embedding
   3. Spectral Normalization
   4. Previous data to the discriminator
   5. MLP discriminator
   6. Embeddings with previous data plus spectral normalization
7. Conclusions
8. Further work
9. References