Add Kokoro TTS plugin (server-mode, via Kokoro-FastAPI)

[HenryVarro666]

Adds livekit-plugins-kokoro: a native TTS plugin for the open-weight Kokoro model served by Kokoro-FastAPI. Fixes #6184.

Why a native plugin (root cause of #6184)

The openai.TTS adapter selects its stream parser by model name: only tts-1/tts-1-hd use the raw-audio path; everything else — including the default gpt-4o-mini-tts and the natural choice model="kokoro" — sends stream_format="sse" and parses the response with iter_lines(), only pushing audio from data: {"type": "speech.audio.delta", ...} events. Kokoro-FastAPI ignores the unknown stream_format field (pydantic extra=ignore) and streams raw binary, so no line ever matches, zero frames are pushed, and the framework raises APIError: no audio frames were pushed — exactly what #6184 reports. Reproduced against a live Kokoro-FastAPI v0.6.0:

[openai-compat default] APIError: no audio frames were pushed for text: ...   (after 3 retries)
[kokoro native]         44 frames, 7.71s audio, TTFB 315ms, total 531ms

The plugin requests response_format="pcm", stream=true (raw 24 kHz mono s16le — Kokoro-FastAPI's sample rate is fixed server-side) and pushes chunks through AudioEmitter with mime_type="audio/pcm" as they arrive.

Relation to #1615

#1615 (@kailashsp) runs Kokoro in-process via kokoro-onnx, and was parked on the shared-inference-process direction davidzhao mentioned there. This PR is complementary, not a replacement: it is server-mode only — an HTTP client with zero in-process inference — so that architectural direction doesn't apply. It follows the same shape as the other HTTP vendor TTS plugins.

Features

• Voice selection incl. Kokoro blending syntax ("af_bella(2)+af_sky(1)"), speed, lang_code
• list_voices() helper (handles both the current and legacy response shapes)
• update_options(), optional bearer token for deployments behind an authenticating proxy (Kokoro-FastAPI itself needs none), KOKORO_BASE_URL env var
• streaming=False capability — Kokoro-FastAPI takes complete text, so the framework's StreamAdapter handles sentence streaming

Testing

Hermetic tests in tests/test_plugin_kokoro_tts.py (pytest --plugin kokoro, no real server or key needed — a local aiohttp fake serves the endpoints): PCM passthrough byte-exactness, request-body correctness, update_options/lang_code, HTTP error → APIError mapping, list_voices both shapes. 4/4 pass.

Live verification against ghcr.io/remsky/kokoro-fastapi-gpu:v0.6.0 (H100): audio artifact plays correctly, voice blending produces distinct waveforms, 68 voices listed. To try it locally:

docker run -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-cpu:latest
python -c "..."  # kokoro.TTS(voice="af_heart") — defaults to http://localhost:8880/v1

make lint / ruff format --check / mypy (strict) all pass.

Note on uv.lock: besides the new package entries, the diff picks up a pending re-lock from an earlier dependency change on main (phonic 0.32.5 → 0.32.9, primp/lxml pruned) — any uv lock run reproduces this. Happy to drop it and let you re-lock separately if preferred.

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.