search

kan-bayashi / ParallelWaveGAN

Unofficial Parallel WaveGAN (+ MelGAN & Multi-band MelGAN & HiFi-GAN & StyleMelGAN) with Pytorch
https://kan-bayashi.github.io/ParallelWaveGAN/
MIT License
1.55k stars 340 forks source link

Training datasets detail of LibriTTS pretrained vocoders #403

Closed emesoohc2022 closed 1 year ago

emesoohc2022 commented 1 year ago

Hi there, I am trying to use LibriTTS pretrained models (HiFi-GAN) as a base model to finetune with other datasets. I have found the pretrained models and its config.yaml in README.md. May I ask the details of training dataset? Did the pretrained model use train-other-500 during training? I see that part = clean is the default config in LibriTTS run.sh.

kan-bayashi commented 1 year ago

I think the pretrained model was trained by clean subset since I wanted to use the same condition as the text2mel model in espnet.