search

ming024 / FastSpeech2

An implementation of Microsoft's "FastSpeech 2: Fast and High-Quality End-to-End Text to Speech"
MIT License
1.79k stars 530 forks source link

how to align Mandarin dataset by mfa？ #59

Open yileld opened 3 years ago

yileld commented 3 years ago

the Mandarin lexicon.txt in your project seems different from mfa pretrained model and it will failed to align，saying "There were phones in the dictionary that do not have acoustic models" image

ming024 commented 3 years ago

@yileld If you want to use your other lexicon, just put it in lexicon/, and change the symbols in text/pinyin.py.

yileld commented 3 years ago

@yileld If you want to use your other lexicon, just put it in lexicon/, and change the symbols in text/pinyin.py.

nonono I didnt mean to change lexicon，I just want to know how you get Mandarin textgrid by your lexicon because it seems mismatch mfa pretrained model，do you use "mfa train xxx" command?

ming024 commented 3 years ago

@yileld yeah I trained the MFA model from scratch since I wish to handle the er-hua （ㄦ話） in the Chinese language carefully, while this feature is not provided in the pretrained MFA model.

hertz-pj commented 3 years ago

@yileld yeah I trained the MFA model from scratch since I wish to handle the er-hua （ㄦ話） in the Chinese language carefully, while this feature is not provided in the pretrained MFA model.

Do you have any plan to share MFA model with you trained to handle the er-hua~

yuxiazff commented 1 year ago

@yileld yeah I trained the MFA model from scratch since I wish to handle the er-hua （ㄦ話） in the Chinese language carefully, while this feature is not provided in the pretrained MFA model.

Do you have any plan to share MFA model with you trained to handle the er-hua~

@ming024 Could you share the mfa acoustic model of AISHELL3，thank you very much!