Real-time-voice-cloning: How to make voices sound better?

[Traincraft101]

Whenever I use the toolbox to make a voice, it sounds not even close to being as good as the demo video was. I tried using the LibreSpeech dataset, LJspeech1.1, and I used high quality wave files, yet it still sounds like a person dying of pneumonia.
Here's what I mean.

Demo video: https://www.youtube.com/watch?v=-O_hYhToKoA&list=WL&index=3&t=85s
My results: https://www.youtube.com/watch?v=_gj30DZ4jLg

The only difference I see in the demo video is that @CorentinJ seems to be using a different vocoder called "gen_s_mel_raw". Where would I get this?

[rfwatkins]

Your GLaDOS examples illustrate the importance of trying a number of samples from the same voice. For mysterious reasons, some samples give much better results than others. Really one just needs to keep trying, and usually you'll eventually hit a really decent sample that gives nice synthesis results. Shorter samples often work better than long ones in my experience thus far. Gaps and sibilant white noise happen when the text sample you're trying to synthesize is too short - too long and it can end up as high-speed babble.

Vocoder effects like GLaDOS don't translate though - you'll still just end up with a normalish sounding voice.

[blue-fish]

Whenever I use the toolbox to make a voice, it sounds not even close to being as good as the demo video was.

Closing this as a duplicate of #162 , @Traincraft101 good observation on the pretrained vocoder as gen_s_mel_raw though I am doubtful that the entire performance difference could be attributed to the vocoder.