feat: add Voxtral TTS engine + French prosody preprocessing

Fix F5-TTS crash on HF Space: NoneType ref_path when no audio uploaded

Fix 3 real bugs found in full code audit

Fix Voice Design voice consistency: full RNG determinism for seed

Implement Rolling Reference technique for Fish-Speech voice consistency

Enable torch.compile for Fish-Speech on GPU (2-3x speedup)

Fix: never sys.exit on missing faster_qwen3_tts — server starts regardless

Add engine gating, CPU/HF Space support, multi-lang Kokoro, and expanded UI params

Add global TTS language selector + more params for Chatterbox & F5-TTS

Fix Qwen3 Voice Design redirect bug + Fish audiobook optimizations

Add engine params (speed/CFG/ref_text), fix dropdown colors, add FR/EN toggle

Add 13 Kokoro voices, improve audiobook UX + chapter detection

Boovore: full UI redesign + Fish-Speech engine + GitHub prep

Add XTTS narrator FR mode: server proxy, UI row, launch scripts

Set 0.6B-Base as default model

Increase text and audio generation limits for longer tests

Add audio size limit (10 MB / ~1 min) and restore text length guard

Add 1000-char text limit to prevent CUDA static cache overflow

Fix stale model ref race, add ACTIVE_MODELS env var, hide unavailable modes

Fix permission error for preset assets and PyTorch cache dir in HF Spaces

Make model cache size configurable via MODEL_CACHE_SIZE env var (default 2, Space uses 5)

Cache all loaded models (LRU, max 5) to avoid re-capture on model switch

Serialize generation with asyncio.Lock and show queue position to users

Download preset wav files from GitHub at startup instead of bundling them

Default preloaded model to 1.7B-Base for voice cloning

Reorder demo UI, default to advanced cloning, auto-select first preset

Move preset buttons inline and rename labels

Add preset voice clone refs in demo

Fix ruff issues

fix(demo): unconditionally import and load nano-parakeet

fix(demo): cap generation at 30s (360 tokens at 12 Hz)

feat(demo): auto-transcribe reference audio with nano-parakeet

feat(demo): arcade mode selector, auto-load 1.7B-CustomVoice default

chore: hf space dockerization

feat(demo): custom voices, xvec toggle, recording, theme

ui(demo): simplify metrics, add advanced toggle

cleanup(demo): drop ws/worklet dead code; cache ref audio

fix(demo): ws meta+pcm framing

feat(demo): binary ws audio chunks

feat(demo): cache ref audio for ws

feat(demo): add wav/base64 timing

feat(demo): add websocket streaming

fix(demo): measure voice clone latency on first chunk

feat(demo): show voice clone latency

Add web demo UI with live TTFA/RTF metrics