epidoux/pocket-tts repository overview

⁠Docker image for pocket-tts

OpenAI-compatible text-to-speech HTTP API wrapping kyutai-labs/pocket-tts⁠. The model is loaded once at container startup and kept warm in memory for every request. Model weights are baked into the image at build time (offline at runtime) for a fast, predictable startup.

pre-built with english (alba) and french (estelle, azelma)

Project version uses semver Major.Minor.Patch.

⁠API

⁠GET /health

{"status": "ok", "model_loaded": true, "sample_rate": 24000}

⁠POST /v1/audio/speech

Same request/response contract as kokoro-fastapi⁠'s /v1/audio/speech, so this image is a drop-in replacement for any client already speaking that contract.

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "pocket-tts", "voice": "alba", "input": "Hello, world.", "response_format": "wav"}' \
  -o out.wav

response_format accepts wav, pcm (no extra dependency), or mp3/opus/aac/flac (requires the image to be built with ffmpeg, on by default).

⁠Environment variables