Version 1.12.5 - Fish Speech with Compilation Speed-Up

[FranckyB]

Fish Speech S2 Pro Integration and Features:

Added Fish Speech S2 Pro (4B parameter TTS model) as a new engine in the UI, supporting 80+ languages and 15,000+ inline expression tags for fine-grained speech control. The model is auto-downloaded from HuggingFace, and [tag] syntax is supported and automatically stripped for other engines.

The TTS model manager (tts_manager.py) now provides get_fish_speech() and generate_voice_clone_fish_speech() methods, handling model loading, reference audio encoding, kernel cache status reporting, and batch/paragraph generation with all Fish Speech parameters exposed.

Performance and Stability Enhancements:

Implemented Triton/Inductor GPU kernel compilation with persistent caching in models/.cache for Fish Speech, significantly accelerating repeat generations. Added Windows-specific patches for compilation compatibility and informative first-run cache warnings.

---

This discussion was created from the release Version 1.12.5 - Fish Speech with Compilation Speed-Up.

[ncwyy2013]

My 5060ti 16g vram shows out of memory. Please add ability to use a lesser vram demand version of fish audio model, like Fish Audio NF4. or use --half to reduce the vram requirements. Thank you!

[AlonDan]

I finally had a chance to try it out today, I had some issues with the RTX 5090 installation, but I shared it here:
#96

I just noticed, sadly it's not yet possible to TRAIN or use LoRA for Fish Speech S2 Pro.

---

Will you consider to add this soon?
For now I will try to train with CMD (Gemini helps with the commands), but it will be wonderful to have it as SIMPLE and EASY as I trained within the GUI of Voice Clone Studio when I trained Vibe Voice, which was super fast btw... I'm aware that Fish Speech S2 Pro demands longer datasets, but I'm gonna try it as soon as you'll implement it.

I started to learn how to prepare dataset for it, that's a lot of work manually at the moment because it needs to insert the [TAG] as well on the .txt files, unlike how simple it was in Vibe Voice... so challenging indeed, but worth trying. 💪

I hope to see it soon, this model is so DYNAMIC that it's insane! I'll be happy to share the experience and feedback.