Fish Audio's Drama 3 Lets Developers Direct AI Voice Like a Human Actor

Fish Audio has released a gated preview of Drama 3, a text-to-speech model that interprets natural-language descriptions of tone, pacing, and character. Fish Audio calls it the most controllable TTS model it has shipped.

The preview introduces three production-focused capabilities: changing a voice within a sentence, rendering multiple characters in one generation, and repairing a single word while preserving the surrounding take. These controls could reduce the regeneration and audio splicing required for dialogue-heavy projects.

Direct prose, finer control

Many expressive TTS systems expose fixed markers such as [whisper], [laugh], or (happy). Drama 3 accepts prose descriptions of the intended performance, giving prompts room to combine delivery traits that a fixed tag vocabulary may not cover.

Calling the preview API

Drama 3 uses Fish Audio’s standard TTS endpoint and is selected through the model request header. According to the TTS API docs, drama-3-preview is a preview identifier whose behavior and availability may change. Requests with an omitted or unrecognized model value fall back to s2.1-pro.

The following request selects Drama 3 and writes the generated MP3 to a local file:

curl --request POST \
  --url https://api.fish.audio/v1/tts \
  --header 'Authorization: Bearer <token>' \
  --header 'Content-Type: application/json' \
  --header 'model: drama-3-preview' \
  --data '{
    "text": "She paused, then whispered: it was you all along.",
    "format": "mp3"
  }' \
  --output take.mp3

The published request shape establishes model selection, but Fish Audio has not documented a formal grammar for Drama 3’s natural-language directions. It also remains unclear whether single-word repair uses this endpoint, a separate API operation, or the preview interface. Developers planning automated editing workflows need those details before integrating the feature.

Model compatibility also varies by synthesis mode. Single-speaker generation works across compatible TTS models. Multi-speaker dialogue works with s2-pro, s2.1-pro, s2.1-pro-free, and drama-3-preview; the older s1 model lacks that capability.

Fish Audio distributed preview API keys through requests on its announcement thread rather than a public self-service setting. Preview pricing has not been published, leaving production costs and usage limits unspecified.

Speaker tokens underneath

Drama 3 builds on multi-speaker work from the S2 family. In the fish-speech repository, Fish Audio explains that S2 can process reference audio containing several voices and represent each speaker with a speaker token. A script can then use speaker IDs to assign dialogue within one generation.

Speaker tokens act as reusable voice identifiers inside the model’s context. They allow one request to preserve separate character identities across multiple turns, removing the need to synthesize every speaker independently and assemble the scene afterward.

Fish Audio says the preview also retains S2-style voice cloning. That workflow derives timbre, speaking style, and emotional tendencies from reference clips lasting roughly 10 to 30 seconds, with no additional fine-tuning. Drama 3 adds per-line performance direction to the cloned voice.

Where editing costs fall

Dialogue generation and localized voice production are the clearest use cases because both require consistent characters, varied delivery, and frequent revisions:

• Short dramas, animation, and games with several characters in each scene.
• Audiobooks that shift voices and emotion within a paragraph.
• Podcast and video voiceovers that receive late script changes.
• Localization and dubbing pipelines that reuse character voices across scenes.

Single-word repair may produce the most immediate workflow savings. A mispronounced name can spoil an otherwise usable clip, and full regeneration often changes timing, emphasis, or cadence elsewhere. A localized replacement preserves the accepted performance and limits downstream editing.

Preview gaps that matter

Fish Audio labels Drama 3 as a preview and warns that its behavior and availability may change. The company has not published a benchmark suite, latency measurements, pricing, or a formal comparison with systems such as ElevenLabs v3 and OpenAI’s voice models.

Natural-language direction also creates reproducibility questions. Instructions such as “sarcastic and nervous” leave room for interpretation, so repeated generations may vary even when the wording stays constant. Production evaluation should measure consistency across voices, scripts, languages, and repeated runs.

Multi-speaker generation needs similar testing around speaker leakage, turn assignment, long-scene consistency, and latency. Fish Audio has not released results for those cases, making the preview better suited to evaluation than customer-facing dependencies with strict reliability requirements.

Voice interfaces become prompts

Drama 3 follows a wider shift in speech synthesis from fixed controls toward instruction-following models. Its central technical bet is that free-form direction can cover combinations of tone, character, and pacing that predefined tags cannot enumerate. The preview’s practical value will depend on how reliably those directions reproduce across takes and whether Fish Audio exposes repair, speaker control, and prompting through stable APIs.