XTTS-v2, Fish Speech stumble on Bulgarian

This is a good reality check for anyone assuming local TTS is just a model download away. For Bulgarian specifically, the bottleneck is language coverage, packaging friction, and Windows/CUDA pain more than raw hardware.

• –XTTS-v2 is a poor fit for Bulgarian out of the box; forcing `language="ru"` is a hack that predictably hurts accent quality and voice similarity.
• –Fish Speech looks better on paper for Cyrillic, but the user ran into the part that breaks most local projects: runtime and model-loading issues, especially on Windows.
• –Blackwell + nightly PyTorch is still a rough edge for developers trying to stay on the bleeding edge of local AI audio.
• –Long-form audiobook narration magnifies every weakness in pronunciation, consistency, and install reliability, so small TTS flaws become deal-breakers fast.
• –ElevenLabs already supports Bulgarian, which makes the open-source/local path feel viable mainly for tinkerers, not for someone who just wants a finished audiobook pipeline.