For years, synthetic speech could read a sentence cleanly and still sound as if it never understood why the sentence mattered. Directors feel that gap on the tenth take: the words are right, the performance is not, and regenerating the whole clip to fix one syllable is often the only public tool on offer.
On 13 October 2026, Fish Audio shipped Drama 3 into that gap. The company wants you to remember a funding round and a slogan about emotion. What moved is narrower. You can download a large speech model, write stage directions into the script, and run the same markup through an API header. How much of the pitch you get still depends on money, GPUs, and invite-only features.
Drama 3 takes bracketed direction and word-scoped tags
Drama 3 is a 35B-parameter mixture-of-experts speech model that activates about 3B parameters per token. Fish Audio pairs that backbone with a new codebook multi-token-prediction audio decoder on the tokenizer from its earlier S2 line. The technical report and Hugging Face card describe hybrid attention, long-form chunking, voice cloning, and as many as sixteen speakers in one script.
In practice, you write the way you would talk to a voice actor. A note in square brackets sets mood or pace. Another can drop a sigh, laugh, or pause where it belongs. Paired tags wrap a few words so a whisper does not infect the rest of the line. Language tags are still experimental; phoneme pins remain for stubborn names. The same text works in Fish Audio’s web editor and in POST https://api.fish.audio/v1/tts once the model header is drama-3. Misspell that header, and the request does not fail loudly. It falls back to s2.1-pro, which is an easy way to fool yourself in a bake-off.
Multi-speaker scenes use tags such as <|speaker:0|> with an array of voice IDs. The direction lives inside the script.
Fish Audio posted 35B weights beside an Artificial Analysis Preview lead
The weights appeared on Hugging Face as fishaudio/drama-3 on launch day. Inference code sits in fishaudio/fish-speech and in SGLang-Omni cookbooks. Fish Audio’s research blog puts the model on fish.audio, on iOS and Android, and on the API as of 13 October, with an 80% API discount through 12 November against a blog list price of $50 per million characters.
Artificial Analysis is the cleaner external signal. On Pronunciation Robustness, human reviewers score hard English spans such as brand names, context-dependent readings, and codes. Drama 3 Preview sits at 93.2%, ahead of Eleven v4 at 91.7% and Eleven v4 Turbo at 90.1%. On the Provider Voice arena, reporting around the launch places that Preview near Elo 1243, roughly tenth overall and the strongest open-weights row in the set, while Eleven v4 Turbo still leads near 1330. AA’s labels still say Preview even though Fish Audio is serving a GA drama-3 header, so read those numbers as Preview-measured until a separate GA row appears.
Fish Audio’s technical report is a house comparison. Multilingual error rates against baselines the company ran, including ElevenLabs v4, come out mixed.
Community License free commercial use stops at $500,000 revenue
Drama 3 is open-weight under the Fish Audio Community License v1.0, dated with the launch. Research and non-commercial use are free. Commercial use is free only while the licensee and its affiliates stay under USD 500,000 in annual revenue from all sources. Offering Drama 3 to others as a TTS, voice-cloning, or inference service needs a separate agreement with Fish Audio. Cloning an identifiable person’s voice without consent is banned. Older S2 weights remain under the Fish Audio Research License. This is not open source in the ordinary software sense.
Hosted on Fish Audio, the model is usable today if you can write direction tags and live with take variance. Self-hosting is realistic for labs and smaller commercial teams inside the revenue cap who can spare roughly 48 GB of NVIDIA memory for FP8, according to the install docs. Building a competing hosted TTS product on these weights without paying Fish Audio is what the licence is written to block.
Around the model sits an older company story. In late July, Fish Audio announced a $52 million seed led by Coreline Ventures and Capital Today and cited $21 million ARR. July materials put users above eight million; the October press release put them above fifteen million. Both figures are Fish Audio’s.
Call drama-3 for directed takes; keep Eleven when arena Elo decides
Teams already rendering through Fish Audio’s TTS endpoint have the cheapest test: change the model header to drama-3, keep a s2.1-pro control, and listen for mid-line direction on the lines you actually ship. The migration cost is one field.
Self-hosters who fit the Community License and can carry the FP8 hardware bill should treat the technical report and the SGLang-Omni cookbook as the runbook. The launch video is marketing.
Where Artificial Analysis’s arena Elo is the figure you have to defend, ElevenLabs and the other closed leaders still sit above Drama 3 Preview. The pronunciation result is worth taking seriously; the overall preference gap has not closed.
Interactive revision and audio inpainting are a separate decision. Fish Audio’s FAQ keeps both in small invite-only testing and promises an editing API later. Until that opens, the public path is still to change the brackets and render again.
Invite-only editing and 48 GB GPUs still bound the pitch
“Most emotive and controllable TTS model in the world” is launch language. CEO Rissa Cao’s press line is more careful: “our most emotive and controllable model yet.” Neither claim is an independent ranking. The expressiveness tables in the technical report are Fish Audio’s own runs, and they do not put Drama 3 ahead of every rival on every row.
Self-host teams that skim past the memory note will discover the limit the hard way. About 48 GB for FP8 is a different buyer from someone clicking play in the web app.
Consent is unfinished. Fish Audio says Uncle Roger’s voice in the launch materials is licensed and used with permission. The primary packet includes no public statement from Nigel Ng or published commercial terms for that voice. The Community License forbids consent-less cloning of identifiable people, yet neither the model card nor the technical report describes a watermark on open-weight outputs. A wrong model header can also leave you grading S2.1 Pro while you think you are grading Drama 3.
Render one line before you rewrite the roadmap
Open-weight speech now has a larger directed model you can download. Whether it moves your stack still comes down to a listening test, a licence read, and a GPU bill. Render one bracketed line on drama-3 before you rewrite the roadmap.