Introduce voice stabilization

#9
by dragonoid - opened

What I mean is that if the model generate a voice we like, it would be cool to export it into a file that allows the preproduction of that exact voice with the same performance no matter where you put it. It happened randomly just now, I kept regenerating the same input text with the same voice description and every time it gave me the same voice, no differences. But this voice that it generated isn't good, but still it provided me the push to send this request here.
It would require a function that analyzes the generated voice at runtime, recording all data and frequencies or whatever this model uses, then exports the file to the user so that when they import it, they get the exact voice. Thanks.

Are you setting a manual_seed using torch before generation? That helps a LOT. Doesn't matter what the seed is, as long as it's always the same.

Sign up or log in to comment