A Pytorch1.4 implementation of HiFi-GAN vocoder adapted from the official github repository, with slight modification on Mel-spectrogram preprocessing.
In the original repo, input mel-spectrogram of target audio is extracted base on pytorch.stft on the fly, while in this repo, input mel-spectrogram of target audio is loaded from stored result of data preprocess.
- Prepare your dataset.
- Extract mel-spectrogram from training audio data;
- Divide dataset into training set & validation set;
- (Optional) Implement your own
get_dataset_filelistmethod inmeldataset.pyto handle metadata.
- Specify following parameters and run
train.pyto start trainging:input_training_file: metadata file of training set;input_validation_file: metadata file of validation set;input_mel_dir: directory where input mel-spectrograms are stored;input_wavs_dir: directory where target audio data are stored (in.npy)checkpoint_path: directory to save model & trainging logsconfig: path of configuration file (in JSON, e.g.config.json)- other optional parameters.
- Download a pretrained HiFi-GAN generator model and corresponding JSON configuration file into the same directory.
- Follow the example in
inference.py, construct aVocoderinstance by specifying the path of the checkpoint file of Generator to be loaded. - Waveform now could be generated from input mel-spectrogram with interface
Vocoder.mel2wav.