WaveNet: A Generative Model for Raw Audio

This is the Chainer implementation of WaveNet

Todo:

• [x] Generating audio

• [ ] Local conditioning

• [ ] Global conditioning

• [ ] Training on CSTR VCTK Corpus

Training the network

Requirements

• Chainer 2

• scipy.io.wavfile

Preprocessing

Donwsample your .wav to 16KHz / 8KHz to speed up convergence.

• Aaudacity

Create data directory

Add all .wav files to /train_audio/wav

Hyperparameters

You can edit the hyperparameters of the network in model.py before running train.py, or edit /params/params.json after training starts.

Training

run train.py

Generating audio

run generate.py

Passing --use_faster_wavenet will generate audio faster than original WaveNet.

Listen to a sample generated by WaveNet

music

Implementation