We present an unsupervised non-parallel many-to-many voice conversion (VC) method using a generative adversarial network (GAN) called StarGAN v2. Using a combination of adversarial source classifier loss and perceptual loss, our model significantly outperforms previous VC models. Although our model is trained only with 20 English speakers, it generalizes to a variety of voice conversion tasks, such as any-to-many, cross-lingual, and singing conversion. Using a style encoder, our framework can also convert plain reading speech into stylistic speech, such as emotional and falsetto speech. Subjective and objective evaluation experiments on a non-parallel many-to-many voice conversion task revealed that our model produces natural sounding voices, close to the sound quality of state-of-the-art text-tospeech (TTS) based voice conversion methods without the need for text labels. Moreover, our model is completely convolutional and with a faster-than-real-time vocoder such as Parallel WaveGAN can perform real-time voice conversion.
Paper: https://arxiv.org/abs/2107.10394
Audio samples: https://starganv2-vc.github.io/
*Our work has been awarded INTERSPEECH 2021 Best Paper Award
- Python >= 3.7
- Clone this repository:
git clone https://github.com/yl4579/StarGANv2-VC.git
cd StarGANv2-VC- Install python requirements:
pip install SoundFile torchaudio munch parallel_wavegan torch pydub pyyaml click librosa- Download and extract the VCTK dataset
and use VCTK.ipynb to prepare the data (downsample to 24 kHz etc.). You can also download the dataset we have prepared and unzip it to the
Datafolder, use the providedconfig.ymlto reproduce our models.
python train.py --config_path ./Configs/config.ymlPlease specify the training and validation data in config.yml file. Change num_domains to the number of speakers in the dataset. The data list format needs to be filename.wav|speaker_number, see train_list.txt as an example.
Checkpoints and Tensorboard logs will be saved at log_dir. To speed up training, you may want to make batch_size as large as your GPU RAM can take. However, please note that batch_size = 5 will take around 10G GPU RAM.
Please refer to inference.ipynb for details.
The pretrained StarGANv2 and ParallelWaveGAN on VCTK corpus can be downloaded at StarGANv2 Link and ParallelWaveGAN Link. Please unzip to Models and Vocoder respectivey and run each cell in the notebook.
For Hifi-GAN support, please check this issue.
The pretrained F0 and ASR models are provided under the Utils folder. Both the F0 and ASR models are trained with melspectrograms preprocessed using meldataset.py, and both models are trained on speech data only.
The ASR model is trained on English corpus, but it appears to work when training StarGANv2 models in other languages such as Japanese. The F0 model also appears to work with singing data. For the best performance, however, training your own ASR and F0 models is encouraged for non-English and non-speech data.
You can edit the meldataset.py with your own melspectrogram preprocessing, but the provided pretrained models will no longer work. You will need to train your own ASR and F0 models with the new preprocessing.
The code for training new ASR models is available here and that for training new F0 models is available here.
- clovaai/stargan-v2
- kan-bayashi/ParallelWaveGAN
- tosaka-m/japanese_realtime_tts
- keums/melodyExtraction_JDC
The author would like to thank @tosaka-m for his great repository and valuable discussions.