Self-hosted, vendor-free text-to-speech for CodeVideo. Pure Node (kokoro-js, q8 ONNX — no Python, no Docker required to run), exposing an OpenAI-compatible speech endpoint. Drops the ElevenLabs single-vendor dependency in the render pipeline.
npm install
npm start # listens on :3000, downloads the ~300MB model on first boot
# or: npm run smoke # synthesize once without the server (writes /tmp/codevideo-tts-smoke.{wav,mp3})ffmpeg is required for MP3 output (apt/brew install ffmpeg); without it the service degrades to WAV rather than failing.
GET /health
POST /v1/audio/speech
body: { "input": "text to speak", "voice"?: "af_heart", "response_format"?: "mp3"|"wav", "speed"?: 1 }
-> audio bytes (audio/mpeg or audio/wav). Cached on disk by sha256(model|voice|speed|format|text).
Example:
curl -s -X POST localhost:3000/v1/audio/speech \
-H 'Content-Type: application/json' \
-d '{"input":"Hello from CodeVideo.","response_format":"mp3"}' --output hello.mp3| var | default | purpose |
|---|---|---|
PORT |
3000 |
listen port |
TTS_API_KEY |
(none) | if set, require Authorization: Bearer <key> |
KOKORO_MODEL_ID |
onnx-community/Kokoro-82M-v1.0-ONNX |
HF model |
KOKORO_DTYPE |
q8 |
model quantization (q8/fp32) |
KOKORO_DEVICE |
cpu |
cpu (onnxruntime-node) or wasm |
KOKORO_VOICE |
af_heart |
default voice |
TTS_CACHE_DIR |
./audio-cache |
on-disk audio cache |
HF_HOME |
(default HF cache) | model cache dir — mount a volume in Docker to persist |
Build the image (docker build -t codevideo-tts .) and add it as a service in
codevideo-api/docker-compose.yml with a named volume on HF_HOME so the model
persists. Point the API's audio generation at it via TTS_SERVICE_URL. The API
keeps uploading the returned audio to S3 (the cost being removed is the
ElevenLabs API, not S3); ElevenLabs stays as an optional BYOK provider.
The previous approach — a docker-compose wrapping the Python kokoro-web
service. Kept as a reference/alternative; the pure-Node service above is the
supported path.