Audio Cycle GAN

This repository features a Cycle Generative Adversarial Network able to generate good quality synthesized audio by improving poorly synthesized audio samples.

Audio samples

TODO

Setup

This repository requires python3, numpy 1.17.2, torch 1.0.1.post2 and librosa 0.7.0.

By default the training set should be set in the following directory structure:

• data
  • fr
    • test
      • A
      • B
    • train
      • A
      • B

I recommand using the french CSS dataset as B. Use any frnech synthetised audio dataset as A (I will provide one later).

Usage

You can see the help page and run the model on the fr dataset like this:

python3 audiocyclegan.py --help
python3 audiocyclegan.py # trains and generate audio samples

It will run the training for 200 epochs and create samples every 10 epochs