This repo contains code implementing my graduation thesis for undergraduate degree.
In short, LiFi-GAN is a HiFi-GAN V1 variant, with a lighter V1-C8C8I generator from iSTFTNet and a lighter MFD discriminator from Basis-MelGAN replacing HiFi-GAN's MPD. It also uses STFT loss from Basis-MelGAN and dropped out feature loss as Basis-MelGAN mentioning it can make the network failing to converge. In our testing, this model can have ~50% faster inference speed and ~50% faster training speed, while maintaining a pretty good audio quality.
Disclaimer: This repo is build for testing purpose. The code is not optimized for performance, nor have being well commented and tested. Modifications are rushed in hurry so a lot of temporary solutions involved. It should not being considered production ready.
It should stays same as HiFi-GAN. You may need to upgrade torch buliding against newer cuda version for newer GPU though.
Please refer to Official HiFi-GAN implementation for more info. In short, use this:
python3 train.py --config config_v1.json
Please refer to Official HiFi-GAN implementation for more info. In short, use this:
python3 inference.py --input_wavs_dir=[where to find input wav] --output_dir=[where to output] --checkpoint_file=[checkpoint file]
We also proivide a extension parameter --device=[cuda or cpu] here.
A copy of generator checkpoint can be retrieved here: link, which is trained on LJSpeech dataset with 1M steps.
This repo is based on: