WaveNet Implementation in TensorFlow

This repository is for a TensorFlow implementation of Google’s WaveNet. The original link of WaveNet paper is here. In this repository, only global conditioning was implemented for WaveNet model.


  • Python 3.5
  • librosa 0.6.2 or higher
  • numpy 1.14.5 or higher
  • tensorflow-gpu 1.10.0 or higher

More simpler way to install these prerequisites using python 3.5 virtualenv is that:

>>> virtualenv --system-site-packages -p python3 wavenet-venv
>>> source ./wavenet-venv/bin/activate
>>> pip install -r requirements.txt


The dataset directory should be organized by this structure.

├── p01 (person name, the name doesn't matter)
    ├── txt
        ├── xxx.txt (name doesn't matter)
        ├── yyy.txt
        └── ...
    └── wav
        ├── xxx.wav (name doesn't matter but should be same with corresponding txt file)
        ├── yyy.wav
        └── ...
├── p02
└── ...

To train a model:

>>> python train.py

Also, to generate a audio with a trained model:

>>> python generate.py


I trained WaveNet model without using any global conditions. The trained model just mimics voice from original audio dataset.


Hyungcheol Noh / About Me

Go back to the List of Projects