search

seungwonpark / melgan

MelGAN vocoder (compatible with NVIDIA/tacotron2)
http://swpark.me/melgan/
BSD 3-Clause "New" or "Revised" License
636 stars 116 forks source link

There is some noise in the gap position #41

Closed AppalachianWine closed 4 years ago

AppalachianWine commented 4 years ago

Hello, thank you very much for the good work! I use Chinese datasets for experiments，and I found some noise in the gap position, May I ask if this is the best result？ this is samples syn.zip

peter05010402 commented 4 years ago

sound good! how long inference one wav?

AppalachianWine commented 4 years ago

The above problems should be caused by my insufficient training, but there are still some other issues. And I think the speed of Melgan is extremely fast! I spent 0.0079s synthesizing 2s of audio in my GPU.

peter05010402 commented 4 years ago

it's faster than G&L!

erogol commented 4 years ago

I experience the same thing with LJspeech. There is a sci-fi like noise at the point of silence.

AppalachianWine commented 4 years ago

Yes, it ’s about this kind of noise, and it may not be completely cleaned.

erogol commented 4 years ago

but I only experience it with predicted Spectrograms meaning usgin melgan with Tacotron. Maybe adding some noise in training might correct it.

AppalachianWine commented 4 years ago

That ’s right, now I ’m not sure if it ’s caused by data or a model.Perhaps this is a solution

MorganCZY commented 3 years ago

@AppalachianWine Have you solved the above issue? I trained this melgan with Biaobei dataset and had the same kind of noise in silence segments.