Add 7 celebrity voices: DiCaprio, Nicholson, Pesci, De Niro, Pacino, Gyllenhaal, Johansson

Add POST /tts endpoint — no URL length limits, send full text in body

Revert to one-shot generation. Chunked crossfade didn't fix stuttering.

Chunked TTS: split long text into sentences, generate each, crossfade joins to remove abrupt cuts

Add 588 voices from kyutai/tts-voices. Total: 665 voices.

Add 24 character voices from chandypants/ollie-pocket-tts (78 total)

FIX: use get_state_for_audio_prompt from Nymbo WAV files (actually produces speech)

FIX: load english_2026-04 model WITH voice cloning (default loads without-voice-cloning = noise)

Fix: convert flat v3/v2 safetensors keys to nested dict format for voice state

Fix: handle v3/v2 pre-computed KV cache embeddings (no audio_prompt key)

Add verbose logging for embedding download failures

Fix: pass HF_TOKEN to huggingface_hub for gated model embedding download

Switch to Docker SDK: FastAPI server with direct WAV/OGG output, no Gradio