This is an unofficial pytorch implementation of stylemelgan - https://arxiv.org/pdf/2011.01557.pdf
The repo is mainly https://github.com/descriptinc/melgan-neurips style
PQMF and stft-losses was used from https://github.com/kan-bayashi/ParallelWaveGAN repo
Several parameters are differ from the paper - segment length, window length in RWD
To do:
- softmax gating is not working --> replaced with sigmoid gating
#2. Discrminitor training still not tested
- - Discriminitor from universal melgan produced better results - combination of time and frequency domain discriminators