Many-to-Many Voice Conversion based on Variational Autoencoder

Code repository for paper link

Manh Luong, Viet Anh Tran under review in INTERSPEECH 2021

Dataset:

We use VCTK-Corpus to train and estimate our proposed model, VCTK dataset can be found in this link

Pretrained model:

pretrained model can be downloaded in this link Wavenet Vocoder: link

Requirements:

• Python 3.6 or newer.
• Pytorch 1.4 or newer.
• librosa.
• tensorboardX.
• wavenet_vocoder pip install wavenet_vocoder

Prepare data for training

1. Download and uncompress VCTK dataset.
2. Move extracted dataset in [home directory].
3. run command: export HOME=[home directory]
4. run command: bash preprocessing.sh.

Usage

To train the model run the following command: bash training.sh

To convert voice from source to target using pretrained model. Run the follwoing commands:

1. cd [Disentangled-VAE directory]
2. mkdir ./results/checkpoints
3. cp [your downloaded checkpoint] ./results/checkpoints/
4. Download pretrained model of Wavenet_vocoder
5. cp [downloaded Wavenet_Vocoder]/checkpoint_step001000000_ema.pth [Disentangled-VAE directory]
6. edit two variables: src_spk and trg_spk in file conversion.sh to your source and target speaker, respectively.
7. run command: bash conversion.sh