Conversation
- copy tokenizer - create clone class of string delimiter but wide (don't judge me for copying code) - remove boost from orig code - add utf8 lib
|
Good afternoon, I have looked through your PR, and it looks fantastic! What is the architectural difference between DE-VITS and EVO VITS? Also, what changed your mind about VITS in the first place? Additionally, is https://github.com/ZDisket/supervits the implementation for EVO VITS, or will it be a separate repository in the near future? |
VITS EVO is basically the successor to that model. It's got much better upgrades, and doesn't depend on external models for improving prosody
I worked on a modern TTS called Echolancer by the end of 2025. That model is excellent at expressiveness, but it's slow. In contrast, this VITS achieves 280x realtime on a single RTX A6000, while still having good quality. I'll be supporting that model too in this, so my plan is VITS EVO for when speed is a priority, and Echolancer when expressiveness is. I'll be doing finetuning as a service.
It's something else. Worry not about it. |
Instead of DE-VITS, time to integrate VITS EVO, with the ONNX DirectML backend.