Implements HiFiGan Neural Vocoder model via PyTorch. Wandb report.
- hw_nv/ is main module of this project; it directly implements HiFiGan Neural Vocoder.
- requirements.txt lists all python requirements.
- setup.sh script installs needed requirements and downloads data and model checkpoint.
- train.py and synthesize.py scripts are used to train model and synthesize audio from mel spectrograms respectively.
To install all requirements and download needed data and last checkpoint simply run:
./setup.sh
If you only need to install python requirements run:
pip install -r requirements.txt
To train model run
python ./train.py -c ./hw_nv/configs/config_V1.json
To synthesise audio run:
python3 ./synthesize.py -c ./hw_nv/configs/config_V1.json -w ./checkpoints/checkpoint_final.pth.tar
Mel spectrograms for synthesis should be placed in directory ./data/validation_mels.
By default setup.sh downloads 3 mel spectrograms and corresponding source audio, and places them in this directory.