StarGAN-Voice-Conversion-2

This is a pytorch implementation of the paper: StarGAN-VC2:Rethinking Conditional Methods for StarGAN-Based Voice Conversion.

**The converted voice examples are in converted directory. The sound quality will improve if more training epochs (~200000) are done.

**VCTK database has been used to train the model with 70 speakers. The convereted samples are a bit noisy because of VCTK data but it can be improved if other clean databases are used.

**Omitted PS in Generator network.

[Dependencies]

• Python 3.5+
• pytorch 0.4.0+
• librosa
• pyworld
• tensorboardX
• scikit-learn
• tqdm

[Usage]

Download dataset

Download and unzip VCTK corpus to designated directories.

mkdir ./data
wget https://datashare.is.ed.ac.uk/bitstream/handle/10283/2651/VCTK-Corpus.zip?sequence=2&isAllowed=y
unzip VCTK-Corpus.zip -d ./data

If the downloaded VCTK is in tar.gz, run this:

tar -xzvf VCTK-Corpus.tar.gz -C ./data

The data directory now looks like this:

data
├── vctk
│   ├── p225
│   ├── p226
│   ├── ...
│   └── p360

Preprocess

Extract features (mcep, f0, ap) from each speech clip. The features are stored as npy files. We also calculate the statistical characteristics for each speaker.

python preprocess.py

This process may take minutes !

The data directory now looks like this:

data
├── vctk (48kHz data)
│   ├── p225
│   ├── p226
│   ├── ...
│   └── p360 
├── vctk_16 (16kHz data)
│   ├── p225
│   ├── p226
│   ├── ...
│   └── p360
├── mc
│   ├── train
│   ├── test

Train

python main.py

Convert

convert.py --src_spk p262 --trg_spk p272 --resume_iters 210000

[Network structure]

[Reference]

• StarGAN-VC2 paper

• StarGAN paper

• CycleGAN-VC paper

[Acknowlegements]

https://github.com/liusongxiang/StarGAN-Voice-Conversion

https://github.com/SamuelBroughton/StarGAN-Voice-Conversion-2