Augmenting Conditional Music Generation

About The Project

Recent advances in deep music generation models are often trained on large datasets containing thousands of hours of music. Due to the large amounts of readily available training data, many contemporary models such as MetaAI's MusicGen forego the implementation of data augmentation during pre-processing. This raises questions regarding the robustness and flexibility of these models against modified input data. This study investigates the performance of MusicGen when exposed to data with the following augmentations:

• Pitch Alterations sampled from a uniform distribution within [-6, 6] semitones
• Volume Adjustments sampled from a uniform distribution spanning [-30, 30] on the midi scale between 0-127
• Tempo Modifications sampled from a discrete distribution comprisiing values of [0.25, 0.5, 0.75, 1.25, 1.5, 1.75]

Getting Started

To run the code yourself,

1. Run Generate_Data.ipynb to generate a file called augmented.zip. The file contains music samples with data augmentation applied and corresponding conditioned outputs generated by MusicGen.
2. Evaluate the quality of the generated samples with the augmented inputs using MFCC.ipynb and RMSE.ipynb.

License

Distributed under the MIT License. See LICENSE.txt for more information.