Fish Audio · 2026-10-13 · major
Fish Audio Drama 3 — an open 35B speech model for multi-character scenes
Fish Audio released Drama 3, an open-weight 35B MoE text-to-speech model (3B active) that voices up to 16 characters, follows inline direction like whispers or laughs, and supports 83 languages. It tops Artificial Analysis pronunciation tests at 93.2%.
Fish Audio's open speech model turns a script with stage directions into a multi-voice performance.
Key specs
| Pronunciation (aa) | 93.2% |
|---|
Quick facts
| Maker | Fish Audio |
|---|---|
| Parameters | 35B MoE, ~3B active per token |
| Languages | 83 |
| Speakers per text | Up to 16 (speaker tags) |
| License | Fish Audio Community License v1.0 |
| Availability | Open weights, fish.audio app, API |
| Serving | SGLang-Omni |
Benchmarks
| Fish Audio Drama 3 | 93.2% | |
|---|---|---|
| Fish Audio S2.1 Pro | 88.1% |
Pricing
| API (launch price) · Oct 13 – Nov 12, 2026 | $10 / 1M characters |
|---|---|
| API (list price) | $50 / 1M characters |
What is it?
Drama 3 voices whole scenes: one text can hold up to 16 characters, and bracketed directions such as a whisper, a laugh or a mid-line change of tone land where the script puts them. Fish Audio's new text-to-speech model also clones voices, handles long-form and streaming audio, and works in 83 languages. The weights are open on Hugging Face.
How does it work?
The backbone is a 35B mixture-of-experts language model (about 3B active per token) initialized from Qwen3.5-35B-A3B, mixing Gated DeltaNet and full-attention layers. A new codebook-MTP decoder predicts the acoustic codebooks for each audio frame, and reinforcement learning uses one learned human-preference reward instead of rule-based scores, so laughs and whispers are not punished for being hard to transcribe.
Why does it matter?
Audiobooks, games and dubbing need acted, multi-voice speech, not one flat narrator. An open model that does this lets small teams run it on their own GPUs, and the license is free for commercial use below US$500,000 in yearly revenue.
Who is it for?
audiobook, game and dubbing developers
Frequently asked questions
- How much does the Fish Audio Drama 3 API cost?
- Fish Audio Drama 3 costs $10 per million characters on the Fish Audio API from October 13 to November 12, 2026, an 80% discount from its $50 list price. In the fish.audio app, Drama 3 uses the same credits as S2.1 Pro, and new Plus, Pro or Max plans get 50% off until November 12.
- Can I use Fish Audio Drama 3 commercially?
- Fish Audio Drama 3 weights are under the Fish Audio Community License v1.0. Research and non-commercial use are free, and commercial use is free for companies under US$500,000 in yearly revenue. Larger companies, or anyone offering Drama 3 as a TTS or voice-cloning service, need a license from Fish Audio. Cloning a real person's voice without consent is forbidden.
- How does Drama 3 compare to Fish Audio S2.1 Pro?
- Fish Audio says Drama 3 is more emotive than S2.1 Pro and follows direction like whispers or laughter more closely. On the Artificial Analysis pronunciation test, Drama 3 scores 93.2% overall against 88.1% for S2.1 Pro. On one H200 it decodes 271 frames per second, versus 177 for S2, despite being over seven times larger.
- How do I run Fish Audio Drama 3 myself?
- Fish Audio Drama 3 weights are on Hugging Face at fishaudio/drama-3 and the code is in the fish-speech GitHub repo. Fish Audio serves the model through SGLang-Omni, which has a Drama 3 cookbook. API users keep the existing S2.1 Pro endpoint and set the model header to drama-3.
Try it
huggingface.co/fishaudio/drama-3