Unofficial implemetation of the RefineGAN vocoder
The RefineGAN vocoder sounds great, let's try to reproduce it!
ℹ it seems to be highly inspired by HiFiGAN, UnivNet & UNet 🎉
Model | param_cnt | train b/s |
---|---|---|
HiFiGAN-v1 | 13936130 | |
RefineGAN | 18962066 |
⚪ install
conda create -n refinegan & conda activate refinegan
- install pytorch follow the official guide
pip install requirements.txt
⚪ inference
python infer.py pretrained/UNIVERSAL_V1/g_02500000
python infer.py pretrained/LJ_V3/generator_v3
⚪ train (refine)
train HiFiGAN:
- download the HiFiGAN repo provided pretrained checkpoints
python train.py -c configs/config_v1.refine.json --load pretrained/UNIVERSAL_V1 --log_path log/test
train RefineGAN:
python train_refinegan.py -c configs/refinegan.json --log_path log/refinegan
Thanks to the code base & pretrained ckpt from official HiFiGAN 🎉~
- thesis
- MelGAN: https://arxiv.org/abs/1910.06711
- HiFi-GAN: https://arxiv.org/abs/2010.05646
- RefineGAN: https://arxiv.org/abs/2111.00962
- repo
- HiFiGAN (official): https://github.com/jik876/hifi-gan
- UnivNet (unofficial): https://github.com/maum-ai/univnet
- pretrained: univ_c16_0292 / univ_c32_0288.pt
- MelGAN: https://github.com/jaywalnut310/MelGAN-Pytorch
- RefineGAN (unofficial): https://github.com/nikhilpinnaparaju/RefineGAN
- RetuneGAN: https://github.com/Kahsolt/TransTacoS-RetuneGAN
by Armit 2023/11/26