Fish Audio Streaming TTS API - Timestamped TTS

Stream synthesized speech with word-level timestamp alignment snapshots for captions, karaoke, and timeline editing.

Stream text-to-speech with word/phrase-level timestamp alignment for karaoke highlighting, synced captions, and timeline editing.

The response is text/event-stream. Each SSE message contains a Base64 audio chunk and, when available, the latest cumulative alignment snapshot for that chunk_seq. Concatenate audio_base64 in arrival order. A newer snapshot for the same chunk_seq should replace the old snapshot — do not append segments.

Base URL

https://api.autorouter.top — Production

{
  "text": "Hello! Welcome to Fish Audio.",
  "reference_id": "model-id",
  "format": "opus",
  "latency": "balanced"
}

Properties:

Name	Type	Required	Description
`text`	string	Yes	Text to synthesize. For multi-speaker, switch with `<\|speaker:0\|>`, `<\|speaker:1\|>`
`reference_id`	string \| string[] \| null	No	Voice ID. A string for a single speaker; an array of IDs for multiple speakers
`references`	object[]	No	Inline reference audio for zero-shot cloning. Requires MessagePack; JSON cannot carry binary
`format`	string	No	Output format: `wav`, `pcm`, `mp3` (default), `opus`. Recommended: `opus` @ 48 kHz
`latency`	string	No	Latency vs quality trade-off: `normal` (default, best quality), `balanced`, `low`. With `balanced`, long text may be split into multiple text chunks
`prosody`	object	No	Speed and volume
`chunk_length`	integer	No	Text chunk length, `[100, 300]`. Default `300`
`normalize`	boolean	No	Whether to normalize Chinese/English text (improves number reading stability). Default `true`
`sample_rate`	integer \| null	No	Sample rate (Hz). When `null`, format defaults apply: 44100 for most formats, 48000 for opus
`mp3_bitrate`	integer	No	MP3 bitrate (kbps): `64`, `128` (default), `192`
`opus_bitrate`	integer	No	Opus bitrate (bps). `-1000` (default, auto), `24000`, `32000`, `48000`, `64000`
`temperature`	number	No	Expressiveness, `[0, 1]`. Default `0.7`. Higher is more diverse
`top_p`	number	No	Nucleus sampling, `[0, 1]`. Default `0.7`
`max_new_tokens`	integer	No	Maximum audio tokens generated per text chunk. Default `1024`
`repetition_penalty`	number	No	Repetition penalty. Values greater than `1.0` reduce repetition. Default `1.2`
`min_chunk_length`	integer	No	Minimum characters before splitting a new chunk, `[0, 100]`. Default `50`
`condition_on_previous_chunks`	boolean	No	Whether to use previous audio as context to keep the voice consistent. Default `true`
`pronunciation_dictionary`	object[]	No	Up to 3 pronunciation dictionaries (platform reference or inline; mutually exclusive)

prosodyobject(Optional)

Speed and volume adjustments.

speednumber(Optional)

Speed multiplier, [0.5, 2.0]. 1.0 is normal speed.

volumenumber(Optional)

Volume adjustment (dB). 0 is unchanged; positive is louder; negative is quieter.

normalize_loudnessboolean(Optional)

Whether to normalize loudness. Supported only by s2-pro and the S2.1-Pro series.

Responses

200: SSE stream

Content-Type: text/event-stream

401: Unauthorized

Content-Type: application/json

402: Insufficient balance

Content-Type: application/json

503: Service overloaded

Content-Type: application/json

Request Example

curl --no-buffer -X POST "https://api.autorouter.top/fishaudio/v1/tts/stream/with-timestamp" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-xxxxxx" \
  -H "model: s2-pro" \
  -d '{
  "text": "Hello! Welcome to Fish Audio.",
  "reference_id": "model-id",
  "format": "opus",
  "latency": "balanced"
}'

Response Example

{
  "audio_base64": "SUQzBAAAAAAA...",
  "content": "Hello world",
  "chunk_seq": 0,
  "chunk_audio_offset_sec": 0.0,
  "alignment": {
    "audio_duration": 0.86,
    "segments": [
      {
        "text": "Hello",
        "start": 0,
        "end": 0.42
      },
      {
        "text": "world",
        "start": 0.42,
        "end": 0.86
      }
    ]
  }
}

Response fields:

Name	Type	Description
`audio_base64`	string	One Base64 audio chunk. Concatenate in arrival order to reconstruct the full audio. This is a transport chunk, not a sentence/word boundary
`content`	string	Text described by the latest alignment snapshot in this event. Long input may be split into multiple content chunks
`alignment`	object \| null	Latest cumulative timestamp snapshot for this `chunk_seq`. When present, replace the old snapshot; do not append. `null` before the first snapshot or when unavailable
`alignment.audio_duration`	number	Duration of the audio generated for this text chunk, in seconds
`alignment.segments`	object[]	Ordered text time segments
`alignment.segments[].text`	string	Segment text
`alignment.segments[].start`	number	Start time in seconds relative to this text chunk's audio start
`alignment.segments[].end`	number	End time in seconds relative to this text chunk's audio start
`chunk_seq`	integer	Text chunk index that the alignment belongs to. Bucket snapshots by this value
`chunk_audio_offset_sec`	number	Absolute start of this text chunk in the full audio, in seconds. Add segment `start`/`end` to get the global timeline

Processing guidance:

Concatenate every audio_base64 in order even when alignment is null.
Store alignments in a map keyed by chunk_seq; each non-null alignment replaces that key.
Add chunk_audio_offset_sec to each segment's local time to get global time.