Fish Audio’s Drama 3 turns stage directions into speech

Fish Audio has released Drama 3, a 35-billion-parameter mixture-of-experts text-to-speech model built for expressive, multi-character audio. Inline cues such as [sighs], [choked up, crying quietly], and [whisper] control the performance from within the script, giving developers a direct interface for tone, timing, and delivery.

At launch, Artificial Analysis ranked Drama 3 Preview as the leading open-weight model in its Speech Arena and first overall in Pronunciation Robustness. The latter result places it ahead of ElevenLabs’ Eleven v4 and Eleven v4 Turbo on difficult words, numbers, and codes, which are common failure points in voice agents, narration, and dubbing systems.

Stage directions become controls

Drama 3 accepts three forms of performance direction, including cues written in a different language from the spoken line:

• Point cues apply an action or delivery change at a specific position in the script.
• Scoped cues apply direction to a selected span of text.
• Global direction describes the mood or performance for a broader passage.

Scripts can assign lines to as many as 16 characters with tags from <|speaker:0|> through <|speaker:15|>. A single request can render an entire conversation while preserving each referenced voice across turns.

Audio inpainting supports targeted corrections without regenerating a full take. To replace a word or short passage, the system receives the earlier take’s vector-quantized audio tokens as context, similar to the prefix supplied when a language model continues existing text. That context helps preserve the surrounding voice, mood, pacing, and rhythm.

A minimal request combines the directed script with a reference recording used to establish the voice:

{
  "text": "[sighs] Fine. If we miss it, it's on you. [laughs] Well, maybe that's fair.",
  "references": [
    {
      "audio_path": "...en0.wav",
      "text": "We asked over twenty different people..."
    }
  ]
}

Sparse compute behind 35 billion parameters