Docoupling voice model generation from text generation.

[perkel666]

The issue.

If I understand it right tortoise does this:

• takes generic model
• finetunes it on .wav files
• generate voice from text based on that finetuned model

Which means each time to produce one sentence it does each time finetuning.

---

---

The solution

• Decouple voice finetuning with .wav files from generation of voice based on text.
• Make script to finetune model with .wavs and save it for future use without generation part.
• Provide a console script to generate voice from text based on finetuned model previously without finetuning it again.

[hesz94]

That's precisely what "get_conditioning_latents" script does. The model isn't finetuned/trained on the voices, but rather extracts "latents" (think "voice description") as a pre-processing step, and then uses these latents "kind of" like a seed for the speech generation.

Sadly decoupling this process won't be a big time-saver, as in the current iteration generation of latents only takes a couple of seconds, however it's still a step forward.