VoiceConversionGANs

GAN series for voice conversion on VCC2018 dataset

This is a voice conversion repository including cyclegan-vc, stargan-vc, stargan-vc2 and some other variants

This work is still in progress, more GAN models will be included

This work is based on repository stargan-vc, stargan-vc2 and cyclegan-vc

Requirements:

1. Python3
2. PyTorch 0.4.1
3. Pyworld

Models:

1. stgan: stargan-vc1 from https://github.com/liusongxiang/StarGAN-Voice-Conversion
2. stgan2: stargan-vc2 from https://github.com/SamuelBroughton/StarGAN-Voice-Conversion-2
3. stgan1_cin: stargan-vc1 + generator with conditional instance normalization + speaker classifier
4. stgan2_new: stargan-vc2 + patchgan discriminator + only target condition in generator and discriminator + no speaker classifier + gradient penalty
5. stgan2_ls: stargan-vc2 + projection discriminator (as in the paper) + source and target conditions in generator and discriminator + LSGAN adversarial loss
6. cycgan: cyclegan-vc1

Preprocess

./run_pre.sh

Modify according to your own conda env and hyper-params

Train:

./run_train.sh

Modify according to your own conda env and hyper-params

Convert

./run_convert.sh

Objective Evaluation

./run_eval.sh

This evaluation calculate Mel Cepstral Distortion (MCD) and Modulation Spectral Distance (MSD) as in stargan-vc2 paper.

However, this script can not get the same score as the paper.