Switch TTS Backend

protoVoice ships two backends: Fish Audio S2-Pro (default) and Kokoro 82M. You pick one via TTS_BACKEND.

Use Fish (default)

TTS_BACKEND=fish docker compose up -d

Fish runs in the fish-speech sidecar container. Requires:

• A second GPU (it's heavy; can't share with routing vLLM).
• FISH_URL reachable from the protovoice container (default http://fish-speech:8092).

See TTS Backends → Fish for tuning.

Use Kokoro

TTS_BACKEND=kokoro docker compose up -d protovoice

Kokoro runs in-process inside the protovoice container. Benefits:

• No sidecar, no second GPU.
• ~50 ms/chunk latency — faster than Fish but less natural.

Pick a voice via KOKORO_VOICE:

TTS_BACKEND=kokoro KOKORO_VOICE=am_michael docker compose up -d protovoice

54 voices are available — see the Kokoro model card for the full list.

Swap at runtime

Not supported yet. Restarting the container with a different TTS_BACKEND takes ~5 s (Whisper + vLLM + Kokoro warm) or ~2 min (Fish cold compile).

Planned for a later milestone: pick a backend per-skill (e.g. Fish for conversational, Kokoro for fast-turn acknowledgements).