Fish Audio S2.1 Pro

Introduction

Most text-to-speech models produce one voice per request. A two-person conversation requires separate API calls for each speaker, then stitching the audio together downstream. S2-Pro handles this differently: you write both sides of the dialogue in a single text input, tag each turn with a speaker index, and the model renders both voices in one audio file with natural turn-taking.

That sample puts two voices and two emotion tags across four speaker turns, all generated in a single request. This guide covers the speaker tag syntax, voice mapping, per-speaker emotion control, and practical patterns for common dialogue scenarios.

Speaker tags

Multi-speaker dialogue uses inline speaker tags to mark where each voice begins. The tags follow the format <|speaker:N|>, where N is a zero-based index that maps to the speech.voices array:

<|speaker:0|>First speaker's line.
<|speaker:1|>Second speaker's line.
<|speaker:0|>First speaker again.

Each tag switches the voice for all text that follows until the next speaker tag. S2-Pro supports two speakers per request (indices 0 and 1). The speech.voices array assigns a voice model ID to each index.

Voice mapping

The speech.voices array maps speaker indices to voice model IDs. Index 0 in the array corresponds to <|speaker:0|> in the text, index 1 to <|speaker:1|>:

[
  {
    "taskType": "audioInference",
    "taskUUID": "b2c3d4e5-f6a7-8901-bcde-f12345678901",
    "model": "fishaudio:s2.1@pro",
    "speech": {
      "text": "<|speaker:0|>The deployment went through without any issues. <|speaker:1|>[excited] Finally! That pipeline has been flaky for weeks.",
      "voices": [
        "536d3a5e000945adb7038665781a4aca",
        "933563129e564b19a115bedd57b7406a"
      ]
    }
  }
]

{
  "data": [
    {
      "taskType": "audioInference",
      "taskUUID": "b2c3d4e5-f6a7-8901-bcde-f12345678901",
      "audioUUID": "a2b3c4d5-e6f7-8901-2345-678901abcdef",
      "audioURL": "https://am.runware.ai/audio/os/a14d18/ws/2/ai/a2b3c4d5-e6f7-8901-2345-678901abcdef.mp3"
    }
  ]
}

Single voice vs. two voices

Dialogue text without speaker tags gets read by one voice. The same text with speaker tags and a voices array produces distinct speakers:

The multi-speaker version distinguishes who is saying what. The model handles turn-taking pacing automatically, with natural pauses between speakers.

Emotion tags per speaker

Bracket emotion tags apply to the current speaker only. Each speaker can carry different emotional delivery in the same passage. The full catalog of tags, free-form expressions, and paralanguage cues is covered in the Emotion and expression guide:

<|speaker:0|>[excited] Guess what? We got accepted into the accelerator program!
<|speaker:1|>[surprised] Wait, are you serious? That is incredible news.

The first speaker carries excitement. The second carries surprise. The tags don't bleed across speakers. Each voice performs its own direction independently.

Dialogue patterns

Formatting multi-speaker text

The model doesn't require line breaks between speaker turns. The <|speaker:N|> tags are the only markers it needs. All of these formats produce the same output:

<|speaker:0|>Line one. <|speaker:1|>Line two.

<|speaker:0|>Line one.
<|speaker:1|>Line two.

Both work. Use whichever format makes your source text easier to read. In JSON payloads, everything goes in a single speech.text string, so line breaks are just \n characters.

Tips

1. Pick voices that contrast. Two voices with similar pitch and cadence sound like one person talking to themselves. Pair a lower-register narrator with a higher-register speaker for the clearest separation.

2. Keep turns long enough to establish the voice. A two-word turn doesn't give the model enough context to differentiate the speaker. Aim for at least one full sentence per turn.

3. Use emotion tags on the speaker who needs them. You don't have to tag every turn. If one speaker is neutral and the other is emotional, tag only the emotional speaker. The contrast makes the emotion more noticeable.

4. Place speaker tags at natural dialogue boundaries. A speaker tag in the middle of a sentence produces an unnatural voice switch. Tag at the start of a new sentence or thought, not mid-clause.

5. Start every text with a speaker tag. The model defaults to speaker 0 if no tag is present at the start. Be explicit by opening with <|speaker:0|> to avoid ambiguity.