WaveGAN

A PyTorch reimplementation of WaveGAN (Donahue, et al. 2018).

Setup

This code requires Python 3 and ffmpeg (which can be installed with conda), along with the following packages (which can be installed with conda or pip):

• pytorch
• numpy
• librosa
• pescador (for sampling)

Running

You can train the WaveGAN with audio from a given directory by using train_wavegan.py. To run with default parameters, run:

python train_wavegan.py <audio_dir> <output_dir>

To see the full list of arguments, run python train_wavegan.py -h. Note that the data split for validation and testing is done on a filewise basis. For SLURM users, an example sbatch script is provided in jobs/train-wavegan.sbatch.

wavegan