search

keithito / tacotron

A TensorFlow implementation of Google's Tacotron speech synthesis with pre-trained model (unofficial)
MIT License
2.95k stars 961 forks source link

Waveglow vocoder #249

Open Clouxie opened 5 years ago

Clouxie commented 5 years ago

Hey there, I'm currently working on Polish model based on keithito implementation. I've succesfully frozen graph for inference. I'm grabbing mel's from BiasAdd:0 and then I've implemented Griffin-Lim in C++ ( "Industrial purposes") however it's still too slow. It is posible to use other network as a vocoder now ? I mean, makes an BiasAdd:0 output as an input to waveglow or other vocoder. Maybe you have some other ideas about faster vocoders, thanks .

mrgloom commented 5 years ago

waveglow should be super slow.

Howewer it's interesting which existing vocoders are compatible with this implementation.

alokprasad commented 4 years ago

@Clouxie Try to use lpcnet c++ as vocoder .its quite fast .There are discussion and project using tacotron2 for generating mels( 20 coefficient ) and then lpcnet as vocoder for realtime solution.