For better speech quality and lower latency, upload reference audio via the create model endpoint. This method uses the Fish Audio SDK and provides a more streamlined approach.

​
Using the Fish Audio SDK

First, make sure you have the Fish Audio SDK installed. You can install it from GitHub or PyPI.

​
Example Usage

from fish_audio_sdk import Session, TTSRequest, ReferenceAudio

session = Session("your_api_key")

# Option 1: Using a reference_id
with open("output1.mp3", "wb") as f:
    for chunk in session.tts(TTSRequest(
        reference_id="MODEL_ID_UPLOADED_OR_CHOSEN_FROM_PLAYGROUND",
        text="Hello, world!"
    )):
        f.write(chunk)

This example demonstrates three ways to use the Text-to-Speech API:

  1. Using a reference_id: This option uses a model that you’ve previously uploaded or chosen from the playground. Replace "MODEL_ID_UPLOADED_OR_CHOSEN_FROM_PLAYGROUND" with the actual model ID.

  2. Using reference audio: This option allows you to provide a reference audio file and its corresponding text directly in the request.

  3. Using a specific TTS model: You can specify which model to use with the backend parameter when calling the tts method. Available options include:

    • "speech-1.5"
    • "speech-1.6"
    • "s1"

Make sure to replace "your_api_key" with your actual API key, and adjust the file paths as needed.

​
Raw API Usage

If you prefer to use the raw API instead of the SDK, you can still use the MessagePack API as described below.

​
Endpoint Details

​
Headers

  • authorization: Bearer token authentication with your API key
  • content-type: application/msgpack
  • model (optional): Specify which TTS model to use. Available options include:
    • speech-1.5
    • speech-1.6
    • s1

​
Example Usage

from typing import Annotated, AsyncGenerator, Literal

import httpx
import ormsgpack
from pydantic import AfterValidator, BaseModel, conint


class ServeReferenceAudio(BaseModel):
    audio: bytes
    text: str


class ServeTTSRequest(BaseModel):
    text: str
    chunk_length: Annotated[int, conint(ge=100, le=300, strict=True)] = 200
    # Audio format
    format: Literal["wav", "pcm", "mp3"] = "mp3"
    mp3_bitrate: Literal[64, 128, 192] = 128
    # References audios for in-context learning
    references: list[ServeReferenceAudio] = []
    # Reference id
    # For example, if you want use https://fish.audio/m/7f92f8afb8ec43bf81429cc1c9199cb1/
    # Just pass 7f92f8afb8ec43bf81429cc1c9199cb1
    reference_id: str | None = None
    # Normalize text for en & zh, this increase stability for numbers
    normalize: bool = True
    # Balance mode will reduce latency to 300ms, but may decrease stability
    latency: Literal["normal", "balanced"] = "normal"


request = ServeTTSRequest(
    text="Hello, world!",
    references=[
        ServeReferenceAudio(
            audio=open("lengyue.wav", "rb").read(),
            text="Text in reference AUDIO",
        )
    ],
)

with (
    httpx.Client() as client,
    open("hello.mp3", "wb") as f,
):
    with client.stream(
        "POST",
        "https://api.fish.audio/v1/tts",
        content=ormsgpack.packb(request, option=ormsgpack.OPT_SERIALIZE_PYDANTIC),
        headers={
            "authorization": "Bearer YOUR_API_KEY",
            "content-type": "application/msgpack",
            "model": "speech-1.6",  # Specify which TTS model to use
        },
        timeout=None,
    ) as response:
        for chunk in response.iter_bytes():
            f.write(chunk)