This is a pytorch implementation of the paper: StarGAN-VC: Non-parallel many-to-many voice conversion with star generative adversarial networks and Non-parallel Voice Conversion using Weighted Generative Adversarial Networks.
**The converted voice examples are in converted directory.
**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.
- Python 3.5+
- pytorch 0.4.0+
- librosa
- pyworld
- tensorboardX
- scikit-learn
- tqdm
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 ./dataIf the downloaded VCTK is in tar.gz, run this:
tar -xzvf VCTK-Corpus.tar.gz -C ./dataThe data directory now looks like this:
data
├── vctk
│ ├── p225
│ ├── p226
│ ├── ...
│ └── p360
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
python main.py
convert.py --src_spk p262 --trg_spk p272 --resume_iters 210000
Note:
- This implementation follows the original StarGAN-VC paper’s network structure, while StarGAN-VC code use StarGAN's network architecture.
- Our converted sound qualities are better than StarGAN-VC code which uses original StarGAN-VC paper’s network structure.
StarGAN-VC code (Original Network Architecture)