Home

Awesome

PPG-GradVC

An any-to-many voice conversion model based on the architecture of Grad-TTS and PPG from a SSL-based phoneme recognizer.

Setup

Using python>=3.6, python<=3.9:

pip install -r requirements.txt

Training

  1. Prepare your multilingual corpora, then fill filelists (lines are formatted in <wavfile_path>|<speaker_num>|.)

  2. A pretrained HiFi-GAN vocoder is located at ./hifigan/g_00875000, you can continue from it or train a new one.

  3. Extract PPGs

    python preprocess_ppg.py --sr 16000 --in_dir /your/dataset/flacs --out_dir /your/dataset/ppgs
    
  4. Set parameters defined in params.py

  5. Run train_ppg.py

Inference

python inference.py -f infer_for_test.txt -c ./logs/grad_300.pt -t 100

Acknowledgement