Fastspeech2 model bad with shorter vocab size ?

TensorSpeech / TensorFlowTTS

:stuck_out_tongue_closed_eyes: TensorFlowTTS: Real-Time State-of-the-art Speech Synthesis for Tensorflow 2 (supported including English, French, Korean, Chinese, German and Easy to adapt for other languages)

https://tensorspeech.github.io/TensorFlowTTS/

Apache License 2.0

3.84k stars 815 forks source link

Fastspeech2 model bad with shorter vocab size ? #180

Closed ZhaoZeqing closed 4 years ago

ZhaoZeqing commented 4 years ago

I run preprocessor with my own symbols which is shorter than LJSpeech symbols, and trained Fastspeech2 model with this new symbols, but the audios synthesized by this model are not good. If I didn't change this line, just use LJSpeech symbols' length, the audios are great. A little puzzled about this.

dathudeptrai commented 4 years ago

:))))))))))))) @ZhaoZeqing you make me puzzle also :)). BTW, chinese will be support tonight, or tmr :))). A quality is great :)).

ZhaoZeqing commented 4 years ago

@dathudeptrai Thanks! I tried to get duration from 15w steps tacotron2, then both ljspeech symbols' length or my own symbols' length are great :) Maybe the model I used to get duration is not so good before ?

dathudeptrai commented 4 years ago

close because inactivity, fell free to re-open or create new issue.