Fish Audio Drama 3
Model iddrama-3
Price$2 per 1M UTF-8 bytes
Checked13 October 2026
Where it stands
No newer speech and audio model in our registry yet.
- Before it from Fish Audio
- Fish Audio S2 Pro / S2.1 Pro9 Mar 2026
- Weights
- Weights are public. Licence: Fish Audio Community License v1.0. See the licence for commercial terms.
How to call it
7 ways to use Fish Audio Drama 3, with the model id each provider expects.
| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Fish Audio API | drama-3 |
https://api.fish.audio/v1/tts (model passed in `model` header; a missing or misspelled header falls back to s2.1-pro) |
docs.fish.audio |
| Fish Audio API (preview build) | drama-3-preview |
https://api.fish.audio/v1/tts |
docs.fish.audio |
| OpenRouter | fish-audio/drama-3 |
https://openrouter.ai/api/v1/audio/speech |
openrouter.ai |
| Hugging Face | fishaudio/drama-3 |
huggingface.co | |
| GitHub | github.com | ||
| SGLang-Omni (self-hosting) | sgl-project.github.io | ||
| Web app | fish.audio |
Pricing
| Per 1M UTF-8 bytes | $2 |
|---|
As published: “USD per 1M UTF-8 bytes on OpenRouter (2026-10-13). Fish Audio’s own pricing page did not list Drama 3 yet; its press release says API usage is 80% off for the first 30 days, so this may be a launch price (list price unverified)”.
Source: openrouter.ai, checked 13 October 2026.
Specs
- Input
- Text and audio
- Output
- Audio
- Context window
- Not published
- Max output
- Not published
- Open weights
- Yes
- Licence
- Fish Audio Community License v1.0 (research and non-commercial free; commercial use only under USD 500,000 annual revenue; no hosted services without a separate agreement)
- Released
- 13 October 2026
- Status
- Current
- Type
- Speech and audio
What stands out
What its maker marketed at launch, and what people found later, each with its source.
Plain-language direction with word-level tags
Free-text direction in square brackets (any language), point tags such as [sighs], and paired tags <whisper>, <emphasis>, <soft>, <fast>, <slow>, <stress>, <forceful> that apply to specific words.
Editing by infilling
Replaces a single word in an existing take while keeping the surrounding audio; vendor-reported Ming-Freeform WER 3.49 (English) and 3.69 (Chinese) against 4.58 and 8.64 for Cartesia Sonic-3 infill.
Multi-turn revision
A note in the next turn (e.g. ‘gentler, and a bit sadder’) regenerates the previous take with it in context; vendor-reported style accuracy 76.5 → 81.8 over three rounds on Step-Audio-Edit.
Multi-character scenes
Speaker tags <|speaker:0|> to <|speaker:15|>: up to 16 characters in one text, with voice cloning from references; long text is chunked and streamed.
Large open-weights speech model with fast serving
35B-parameter MoE backbone, 3B active per token; vendor-reported 88.9 ms median time to first audio and RTF 0.131 for one stream on one NVIDIA H200 with SGLang-Omni.
Notes
Preview since 2026-09-23 (drama-3-preview). Weights about 39.6 GB plus a 1.87 GB codec; about 80 languages (the HF card lists 83 codes). Artificial Analysis (2026-10-13): ‘Drama 3 Preview’ #10 overall, Elo 1243, top open-weights model, below Eleven v4 (1328-1330). Benchmarks are vendor-run and mixed: lower error rate than ElevenLabs v4 in 9 of 24 languages, weaker in Japanese, Indonesian, Greek, Romanian. HF card metadata still names the Research License, but the LICENSE file is the Community License. No watermark is documented for the open weights.
Sources: https://huggingface.co/fishaudio/drama-3 , https://huggingface.co/fishaudio/drama-3/blob/main/LICENSE , https://github.com/fishaudio/fish-speech , https://openrouter.ai/fish-audio/drama-3 , https://artificialanalysis.ai/text-to-speech/leaderboard
Quick start (Fish Audio API): POST https://api.fish.audio/v1/tts with header model: drama-3, a bearer API key, and a JSON body with text and a reference_id (voice). Self-hosting: sgl-omni serve --model-path fishaudio/drama-3 --port 8000 (from the model card).