HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis

Intro

This repo is a fork from https://github.com/jik876/hifi-gan, with very few modifications:

• It keeps the old format of the aforementioned repo.
• It is adjusted to read given mel spectrograms.

To understand HiFi-GAN read the original paper.

Pre-requisites

1. Python >= 3.6
2. Clone this repository.
3. Install python requirements. Please refer requirements.txt

Training

python train.py \
  --config configs/config_v1.json \
  --input_wavs_dir dataset/ptBR/audio \
  --input_mels_dir dataset/ptBR/mels \
  --input_training_file dataset/train_files.txt \
  --input_validation_file dataset/test_files.txt

To train V2 or V3 Generator, replace config_v1.json with config_v2.json or config_v3.json.
Checkpoints and copy of the configuration file are saved in cp_hifigan directory by default.
You can change the path by adding --checkpoint_path option.

Pretrained Model

Here is a pretrained model trained on 20hs of a multispeaker ptBR dataset.

Inference for end-to-end speech synthesis

1. Make test_mel_dir directory and copy generated mel-spectrogram files into the directory.
   The spectrograms produced by this model are compatible with the pretrained checkpoint, for instance: Tacotron2.
2. Run the following command.

   python inference_e2e_from_folder.py --input_mels [test_mel_dir] --output_dir --output_dir [output_wav] --checkpoint_file [generator checkpoint file path] --npyin True
   

Generated wav files are saved in output_wav.

Acknowledgements

Many thanks to Jungil Kong, Jaehyeon Kim, Jaekyoung Bae for making the original repo available.