Fish Audio WebSocket TTS API - Real-time Speech Synthesis

Bidirectional streaming TTS over WebSocket. The client pushes text chunks; the server pushes audio chunks.

Low-latency bidirectional speech synthesis over WebSocket. The client streams text; the server streams audio. Messages are serialized with MessagePack.

The request field inside StartEvent matches HTTP TTS parameters. In WebSocket mode, request.text is usually empty; actual text is sent in later TextEvents.

Authentication

BearerAuth: http (bearer) Authenticate using a Bearer Token. Format: Authorization: Bearer sk-xxxxxx

The connection may set the model request header: s1, s2-pro, s2.1-pro, s2.1-pro-free. If omitted or unrecognized, it falls back to s2.1-pro. For multi-speaker dialogue, use s2-pro or the S2.1-Pro series.

Client event order:

After connecting, send start once (with TTS config)
Send text in chunks
Optionally send flush to synthesize buffered text immediately
After all text is sent, send stop

Server events: Continuously return audio binary chunks, then send finish when done (reason is stop or error).

Request Body

Content-Type: application/msgpack

Start example:

{
  "event": "start",
  "request": {
    "text": "",
    "reference_id": "model-id",
    "format": "opus",
    "latency": "balanced"
  }
}

Properties (StartEvent):

Name	Type	Required	Description
`event`	string	Yes	Fixed to `start`
`request`	object	Yes	TTS configuration; fields match HTTP TTS
`request.text`	string	Yes	Usually an empty string in WebSocket mode
`request.reference_id`	string \| string[] \| null	No	Voice ID. A string for a single speaker; an array for multiple speakers
`request.references`	object[]	No	Inline reference audio (zero-shot cloning)
`request.format`	string	No	Output format: `wav`, `pcm`, `mp3`, `opus`
`request.latency`	string	No	`low` / `normal` / `balanced`
`request.prosody`	object	No	Speed and volume
`request.temperature`	number	No	Expressiveness, `[0, 1]`. Default `0.7`
`request.top_p`	number	No	Nucleus sampling. Default `0.7`
`request.max_new_tokens`	integer	No	Maximum audio tokens per text chunk. Default `1024`
`request.min_chunk_length`	integer	No	Minimum characters before splitting a new chunk, `[0, 100]`. Default `50`
`request.chunk_length`	integer	No	Text chunk length
`request.normalize`	boolean	No	Whether to normalize Chinese/English text
`request.sample_rate`	integer \| null	No	Sample rate (Hz)
`request.condition_on_previous_chunks`	boolean	No	Whether to use previous audio to keep the voice consistent

Subsequent client eventsobject

Text and control events.

textobject(Required, repeatable)

{ "event": "text", "text": "Hello! Welcome to Fish Audio." }

One event per text chunk. For multi-speaker, use <|speaker:0|> markers in the text.

flushobject(Optional)

{ "event": "flush" }

Force immediate synthesis of text still in the buffer.

stopobject(Required)

{ "event": "stop" }

Close the session after all text has been sent.

Responses

101: WebSocket established; MessagePack events follow

401: Unauthorized

Content-Type: application/json

Request Example

# WebSocket 无法用普通 curl 完整演示，请使用 wscat 或下方 SDK 示例
wscat -c "wss://api.autorouter.top/fishaudio/v1/tts/live" \
  -H "Authorization: Bearer sk-xxxxxx" \
  -H "model: s2.1-pro"

Use MessagePack encoding and decoding in production, not JSON. The examples above only show event order.

Response Example

{
  "event": "audio",
  "audio": "<binary audio chunk>"
}

Response fields:

Name	Type	Description
`event`	string	`audio` or `finish`
`audio`	binary	Audio chunk. Concatenate in arrival order
`reason`	string	Finish reason. `stop` is a normal end; `error` means a failure