apeklets / kicks-thesis

Bachelor Thesis: Kick Drum Generation with Neural Networks.

Kick Drum Generation with Neural Networks

This repository contains the code used to conduct the experiments for our bachelor thesis. You can read the full text here. We have an accompanying site with sound examples here.

Running experiments

Each of the models has its own code dependencies, so we ran each model in a different Python virtual environment. We used Poetry for this, but any other environment management will work. The approach is to create a virtual environment in the correct folder, install the dependencies listed in pyproject.toml or requirements.txt, and then run pip install -r requirements_cuda110.txt to install the latest PyTorch version in that virtual environment.

Progressive WaveGAN

The folder experiments/wavegan contains our adaptation of the WaveGAN model, based on this PyTorch version.

Other models

We compared the progressive WaveGAN versions to some other models:

• SpecGAN
• Nistal PGAN
• WaveRNN
• WaveNet

These all also got their own folder in experiments/, and we made some small changes to adapt these to our goals.

Quality Measure

The folder notebooks/quality contains the source code for the quality measure we defined for evaluating kick drum sample quality. The folder notebooks/ further contains some jupyter notebooks that document some informal experiments we conducted during the design of this quality measure.

Results

The results folder contains some config file, samples and quality measure graphs from a couple of our experiments.