search

SJTMusicTeam / Muskits

An opensource music processing toolkit
Apache License 2.0
309 stars 44 forks source link

is xiaoice pretrain models trained by world feat? #118

Closed 11721206 closed 2 years ago

11721206 commented 2 years ago

hello， when I analysis code to find world feat not used in code？I want to confirm that wav->sp->mgc(60) has been done in this repo? and the residual connection in xiaoice has been done in this repo? thank you

ftshijt commented 2 years ago

Thanks for your interest in the toolkits. We use log-mel spectrogram features instead of world feature here. The feature is extrated on-the-fly at https://github.com/SJTMusicTeam/Muskits/blob/main/muskit/svs/feats_extract/log_mel_fbank.py

To get a compatible vocoder, you can use vocoder trained at https://github.com/kan-bayashi/ParallelWaveGAN (we have also published several vocoders for our recipes there)

ftshijt commented 2 years ago

Given that we do not condition on explicit pitch information for the vocoder, we do not take the residual connection for the pitch as in original Xiaoice paper