fish-speech-api

A Docker image that runs the Fish Speech S2 Pro TTS API server, with the model weights baked in at build time.

How it works

The image builds on fishaudio/fish-speech:server-cuda, downloads the fishaudio/s2-pro checkpoints during the build, and starts tools/api_server.py on container start. The server listens on port 8080 and exposes the standard Fish Speech HTTP API (/v1/tts, /v1/asr, /v1/health, ...).

Build

docker build -t fish-speech-api .

Run (requires NVIDIA GPU)

docker run --gpus all -p 8080:8080 fish-speech-api

Configuration

Override the defaults with environment variables:

Variable	Default	Description
LISTEN	0.0.0.0:8080	Host and port the API server binds to
DEVICE	cuda	Torch device (cuda, cpu, ...)
LLAMA_CHECKPOINT_PATH	/app/checkpoints/s2-pro	Path to the LLaMA checkpoint directory
DECODER_CHECKPOINT_PATH	/app/checkpoints/s2-pro/codec.pth	Path to the decoder checkpoint

docker run --gpus all -p 9000:9000 -e LISTEN=0.0.0.0:9000 fish-speech-api

Usage

The /v1/tts endpoint accepts a msgpack-encoded ServeTTSRequest and returns the raw audio bytes.

import ormsgpack
import requests

from fish_speech.utils.schema import ServeReferenceAudio, ServeTTSRequest

request = ServeTTSRequest(
    text="Hello, world!",
    references=[],           # or [ServeReferenceAudio(audio=<bytes>, text="...")]
    reference_id=None,
    format="wav",            # wav, mp3, flac
    max_new_tokens=1024,
    chunk_length=300,
    top_p=0.8,
    repetition_penalty=1.1,
    temperature=0.8,
    streaming=False,
    use_memory_cache="off",
    seed=None,
)

response = requests.post(
    "http://127.0.0.1:8080/v1/tts",
    params={"format": "msgpack"},
    data=ormsgpack.packb(request, option=ormsgpack.OPT_SERIALIZE_PYDANTIC),
    headers={"content-type": "application/msgpack"},
)

with open("out.wav", "wb") as f:
    f.write(response.content)

License

MIT

Project structure

.
├── src/
│   └── run.sh       # Startup script
├── checkpoints/     # Downloaded at build time (gitignored)
└── Dockerfile