Add DE-VITS (VITS + TM + BERT)

[ZDisket]

This PR will add DE-VITS (codename VITS iCMOTT "6 Elements"). It will replace VITSTM in-place.

• Create BERT tokenizer
• Create BERT model inference
• Create DE-VITS class
• Finalize implementation

[TheDevloper2023]

When are you going to finish this actualy?

[ZDisket]

When are you going to finish this actualy?

Never. In late 2023 I found that VITS is inadequate for the kind of speech I want, as the duration predictor is incapable of modeling time series accurately. I'm working on a semi-custom TTS model to replace it, ETA is like 6 months to release.

[TheDevloper2023]

When are you going to finish this actualy?

Never. In late 2023 I found that VITS is inadequate for the kind of speech I want, as the duration predictor is incapable of modeling time series accurately. I'm working on a semi-custom TTS model to replace it, ETA is like 6 months to release.

Cool. How will the arhitecture work? And how would I control the outputs emotion?
You should realese an training script for it, that would be nice.
One more question: Do You Need an emotional dataset with labels or not.

(Sorry If my grramar sucks, English is NOT my first languege)