An Easy Implementation of HIFIGAN with muit-band.
This project is based on:
- Offical implementation of HIFI-GAN by jik876
- Unoffical implementation of ParallelWaveGAN by kan-bayashi
Something new
- Support different datasets
- Support muti-band
- Support different sample_rate, hop_size ...
Current Training...
Firstly, the datasets structure like below.
datasets:
wavs:
speaker1:
???.wav
speaker2:
???.wav
.....
Secondly, create training.txt
python preprocess.py
Thirdly, training
python train.py
Some problems
while using pqmf merge four subbands(value between -1 and 1), value may greater than 1.0 and less than -1.0. I think it may cause by PQMF filters (can't reconstruct perfectly)