search

keonlee9420 / Expressive-FastSpeech2

PyTorch Implementation of Non-autoregressive Expressive (emotional, conversational) TTS based on FastSpeech2, supporting English, Korean, and your own languages.
Other
267 stars 46 forks source link

Mechanical noise #8

Closed MrYANG23 closed 2 years ago

MrYANG23 commented 2 years ago

hi nice work, I have a issue，how to resolve Mechanical noise，no matter FastSpeech2 or this repo，there always exiest Mechanical noise

keonlee9420 commented 2 years ago

Hi @MrYANG23 , what do you mean by 'Mechanical noise'?

bagustris commented 2 years ago

Hi @keonlee9420,

Maybe what he meant by mechanical noise is like in the following file, https://www.dropbox.com/s/0jo0largys117cz/That%27s%20out%20of%20control_Ses01M_ang.wav?dl=0

In my case, I only heard the noise only. My setup is the default one in IEMOCAP dataset. Any idea?

keonlee9420 commented 2 years ago

Ah, I see. It is mainly originated from the dataset (especially IEMOCAP). There is a few options from my experience:

  1. the currently shared vocoder is not trained on IEMOCAP, so you can fine-tune it on IEMOCAP and/or model output.
  2. you can also utilize speech enhancement to denoise them. I tried FullSubNet, and it showed somewhat cleared audio.
  3. of course, the best way to this end is preparing super studio-clear quality audio. But I know it is not an easy task in general.
MrYANG23 commented 2 years ago

Ah, I see. It is mainly originated from the dataset (especially IEMOCAP). There is a few options from my experience:

  1. the currently shared vocoder is not trained on IEMOCAP, so you can fine-tune it on IEMOCAP and/or model output.
  2. you can also utilize speech enhancement to denoise them. I tried FullSubNet, and it showed somewhat cleared audio.
  3. of course, the best way to this end is preparing super studio-clear quality audio. But I know it is not an easy task in general.

thank your reply ,yes what I mean like this issue https://github.com/ming024/FastSpeech2/issues/76

keonlee9420 commented 2 years ago

close due to inactivity