Fish Audio releases Drama 3

Fish Audio released Drama 3 on 2026-10-13, three weeks after a preview (API id drama-3-preview, 2026-09-23). The company published the weights on Hugging Face, inference code in fishaudio/fish-speech, a technical report, and a press release. The launch post on X said: “Introducing Drama 3: the most emotive and controllable TTS model in the world. And yes, it’s open-weight.”

The model is much larger than its predecessor S2 Pro: a 35B-parameter mixture-of-experts backbone that activates 3B parameters per token, against S2’s roughly 4.4B. According to the technical report, four functions are new and live in one model:

• Revision from a note: the next turn can say “gentler, and a bit sadder” and the model redoes the take with the earlier one in context.
• Editing by infilling: one word can be replaced (“Wednesday” → “Thursday”) while the audio around it is kept.
• Scene production: several characters (up to 16 speaker tags) in one pass.
• Streaming text input and long-form synthesis.

Direction is written as free text in square brackets, in any language. Paired tags such as <whisper>…</whisper> limit a delivery to specific words, which is what the launch post means by controlling “how even a single word should sound”.

Drama 3 puts features that were mostly limited to closed services (take-by-take direction, single-word repair, multi-character scenes) into downloadable weights, and it is the largest open-weights speech model in this dataset. It overtook Breeze TTS 2 as the top open-weights model on Artificial Analysis while staying below the closed Eleven v4 released two weeks earlier. The revenue-capped licence is a middle path between research-only weights and permissive open source.

The weights are downloadable and not gated, but the licence is restrictive. The Fish Audio Community License v1.0, dated the launch day, allows research and non-commercial use for free and commercial use only “while Your Annual Revenue is less than USD 500,000” (gross revenue of the licensee and affiliates from all sources). Offering the model as a hosted service, an API or a voice-cloning product needs a separate agreement whatever the revenue, and outputs may not be used to train other speech models. Artificial Analysis marks the model as open weights without commercial openness. This is more permissive than S2 Pro, whose weights stay under the non-commercial Fish Audio Research License.

The Hugging Face card is inconsistent: its metadata still names fish-audio-research-license, while the card text and the LICENSE file are the Community License.

Vendor claims (not independently checked):

• “Most emotive and controllable TTS model in the world” (X post); “our most emotive and controllable model yet” (CEO Rissa Cao, press release).
• All benchmark tables in the technical report are Fish Audio’s own runs, including the ElevenLabs v4, Inworld TTS-2, Soniox and Cartesia Sonic-3 baselines. The report itself shows mixed results: Drama 3 has the lowest error rate of all listed systems in Chinese, Dutch, English, German and Turkish, loses to ElevenLabs v4 in most other languages, has a higher error rate than S2 in Japanese, Korean and Indonesian, and scores below Step-Audio-EditX on emotion accuracy and paralinguistic events. The report has no benchmark for expressiveness against ElevenLabs.
• Latency and throughput figures (88.9 ms to first audio on one H200).
• 15M+ users and $21M ARR. In July 2026 the company reported more than 8 million users and the same ARR.

Independently checked on 2026-10-13:

• Weights, file sizes, licence text and the code commit (“Update for Drama-3”, 15:29 UTC) are public.
• Artificial Analysis lists “Fish Audio Drama 3 Preview” at #10 overall with Elo 1243, the best open-weights model, 22 points above Breeze TTS 2 and 87 points below Eleven v4 Turbo (1330) and Eleven v4 (1328). This is the preview build; the released model had no separate row yet.
• OpenRouter serves fish-audio/drama-3 at $2 per 1M UTF-8 bytes.
• A community demo Space (mrfakename/drama-3) appeared on Hugging Face within two hours.

Not found: a Hacker News front-page story, press tests of the released model, or third-party listening comparisons with Eleven v4. A Reddit search could not be completed. Unite.AI noted that the release gives no independent evaluation of how well single-word repairs keep the surrounding audio.

The launch video uses the voice of Uncle Roger, the character of comedian Nigel Ng. Fish Audio’s product page calls it “Uncle Roger’s licensed voice”. We found no post or statement from Ng about it and no public terms, so whether and how he was paid is unverified. Separately, user-uploaded “Uncle Roger” voices exist in Fish Audio’s public voice library.

The licence forbids cloning an identifiable person’s voice without consent and forbids removing “any watermark or provenance signal applied by the Fish Audio Materials”. Neither the model card nor the technical report describes a watermark, so it is unverified that the open weights apply one. In July 2026 TechCrunch reported creators’ complaints that their voices had been uploaded to Fish Audio without consent; the company said takedowns are now automated.