Creating Music Using a Generative Adversarial Network

By Billard, Mitchell; Bishop, Robert; Elsisy, Moustafa; Graves, Laura; Dr. Kolokolova, Antonina; Nagisetty, Vineel; and Northcott, Zachary: all affiliated with Memorial University of Newfoundland

Table of Contents

• Introduction
• Usage
• More Info

Introduction:

Our research aims to discover if successful AI image generative models can also be used to generate music. We plan to use a Deconvolutional Generative Adversarial Network, a particular type of Artificial Neural Network, in an attempt to generate Irish music.

We hope to use the distinctive structure of Irish tunes to make them suitable for GAN-based music generation. Our main idea is to regard an Irish melody as a fixed-size object with cross-references among its parts: “music as a picture” view. The two key components of this project are:

• Preprocessing:
  • Creating a well-defined format for music encoding suitable for GANs
  • Representing “vertical” dependencies in the tunes through a music as a picture view
• Modifying GANs via:
  • Structuring strides and kernel size to take advantage of observed features in image forms
  • Modifying training rate to avoid “Mode Collapse”

Usage:

Source Code:

• The Preprocessing code is found in the src/Generation/ directory
• The Neural Network training code is found in the src/Model/ directory
• The Data used for preprocessing and training is found in the Data/ directory
• The code for our website is found in the Site/ directory

Reproduce Results:

• To reproduce the results shown in the report, please run the notebook file Ensemble.ipynb which is found in the Documentation/ directory.

More Info

• More information is found here.
• The abstract report for this project is found here.