Fish Audio Drama 3: direct the performance, not just the words
Fish Audio's Drama 3 (October 2026) is an open-weight 35B text-to-speech model that voices whole scenes — up to 16 characters in one script, with stage directions like [laughs] or a whisper landing exactly where you write them.
The demo works the reverse direction — speak, get clean text back. Drama 3 goes text → performed speech; unmuss goes speech → text.
What is Drama 3
Released on October 13, 2026, Drama 3 is Fish Audio's new-generation text-to-speech model — and, by total parameters, the company calls it the largest open-weight speech generation model to date [official]. The backbone is a 35-billion-parameter mixture-of-experts language model with about 3B parameters active per token, so per-token compute stays near the previous generation while capacity grows [official].
The headline feature is directing, not just synthesis. One script can hold up to 16 characters via speaker tags, and bracketed directions — a whisper, a laugh, a mid-line change of tone — land exactly where the script puts them [official]. It also clones voices from a short reference clip, handles long-form text in streamed chunks, and works in 83 languages [official].
Key specs
| Item | Detail |
|---|---|
| Model | 35B mixture-of-experts, ~3B active per token [official] |
| Speakers per script | Up to 16, via <|speaker:0|> … <|speaker:15|> tags [official] |
| Direction | Inline tags like [laughs], [sighs], whispers, mid-line tone changes [official] |
| Languages | 83 [official] |
| Voice cloning | From a short reference clip + transcript [official] |
| Weights | Open on Hugging Face, Fish Audio Community License v1.0 [official] |
| API price | $10 / 1M chars launch (until Nov 12, 2026), then $50 / 1M list [reported] |
| Benchmark | 93.2% on Artificial Analysis Pronunciation Robustness — tops the test [reported] |
How to direct it
-
1. Cast your characters — wrap each speaker's lines in speaker tags (
<|speaker:0|>…<|speaker:15|>). One request renders the whole scene, no per-line rendering. -
2. Write stage directions inline —
[laughs],[sighs],[whispers]inside the line, exactly where the moment happens. The model was trained with a learned human-preference reward so expressive moments aren't punished for being hard to transcribe. -
3. Clone or pick voices, then revise in context — attach a reference clip to clone a voice, and send earlier takes as context when re-rendering a line so pacing and tone carry over.
Drama 3 vs single-voice TTS
| Single-voice TTS | Drama 3 | |
|---|---|---|
| Speakers | One voice per request | Up to 16 characters in one script |
| Emotion | Global style or SSML tweaks | Inline direction tags at exact positions |
| Dialogue workflow | Render each line, place by hand | Whole scene in one pass, revise in context |
| Weights | Usually API-only | Open weights, self-hostable (48GB+ GPU) |
Pricing
The API launched at $10 per 1M characters from October 13 to November 12, 2026, reverting to the $50 list price after [reported]. The weights are open under the Fish Audio Community License v1.0 — free commercial use up to $500,000 in annual revenue [reported]. It's also usable with no code in the fish.audio web app and the iOS/Android apps [official].
Who it's for — and who it's not for
For
Podcast dramas, game dialogue, audiobooks with full casts, video narration with directed emotion — anyone who writes scripts with more than one voice in them.
Not for
Going the other direction. If your problem is spoken audio you need as clean text — meetings, interviews, voice memos — that's the speech-to-text side, which is what unmuss does.
Frequently asked questions
- Is Fish Audio Drama 3 free?
- The weights are open under the Community License (free commercial use up to $500k annual revenue). The hosted API launched at $10/1M characters until Nov 12, 2026, then $50/1M list. There's also a free tier in the fish.audio app — check their site for current limits.
- How is Drama 3 different from ElevenLabs?
- Two big differences: Drama 3 voices up to 16 characters in a single script with inline stage directions, and its weights are open for self-hosting. ElevenLabs remains the incumbent for single-voice quality and voice library breadth — pick by whether your script has a cast.
- What hardware do I need to self-host it?
- Per launch coverage, 48GB+ of NVIDIA VRAM in FP8 for the full model. It serves via SGLang-Omni or vLLM-Omni from launch day.
- Can it clone my voice?
- Yes — attach a short reference clip plus its transcript and the model clones the voice for the script. Only clone voices you own or have permission to use.
- How does this relate to unmuss?
- Opposite directions of the same voice pipeline. Drama 3 turns text into performed speech; unmuss turns your speech into structured, cleaned-up text — try the demo at the top of this page. Full policy: /privacy/.
More voice tools
Work with voice in both directions.
Drama 3 performs your scripts. unmuss writes down your thoughts — download it and stop losing ideas to thin air.
Download on the App Store