Fish Audio Drama 3: An Open-Weight AI Voice You Can Direct

Fish Audio released Drama 3 on October 13, 2026. It is a text-to-speech model with a 35-billion-parameter mixture-of-experts backbone, about 3 billion parameters active per token, that accepts stage directions in square brackets and applies them to the words that follow. The weights are public on Hugging Face under the Fish Audio Community License version 1.0, which permits commercial use at no charge while your annual revenue is under US$500,000 and bars any hosted speech service without a separate agreement. The API lists at $50 per million characters, with an 80% launch discount running to November 12.

Two things make this worth a post rather than a line in our October release tracker. The first is the control surface. Most voice models take an emotion label or a style preset; Drama 3 takes a direction written in plain words, in any of its 83 languages, placed exactly where the change should happen. The second is the licence, which is more specific than the usual open-weight boilerplate and draws its line in a place that matters to agencies: a studio producing voiceover for a client is fine, a product that lets customers generate speech is not, whatever the revenue.

Everything below is Fish Audio's own statement unless marked otherwise. The one benchmark figure is a vendor-cited score from a third party. We have not run the model, and this post makes no quality claim of its own.

The launch post is signed by Fish Audio's chief scientist Shijia Liao, chief executive Rissa Cao and chief technology officer Jiahua Liu. It describes three things: a direction system built on square brackets and paired tags, a set of editing features still in invite-only testing, and a release of the weights and code. The model's own page on Hugging Face adds the architecture: a mixture-of-experts text backbone with hybrid full and linear attention, an audio decoder that predicts ten codebooks causally across the sequence, FP8 experts with the rest of the weights in bf16. The launch post names the base: the language model of Qwen3.5-35B-A3B, with 40 layers.

What it is not: a conversational speech-to-speech model, a voice agent runtime, or a music model. It is text in, speech out, with the direction embedded in the text. Our September ranking of text-to-speech models places its predecessors; Drama 3 is new since that table was built and is not in it.

That sentence is the vendor's claim about its own model and we print it as such. The rest of the post is organised around the parts of the release a buyer can check: the control syntax, the licence text, the price page and the one third-party number.

The control model has two parts. A direction in square brackets is a note to the performer, and Fish Audio's description is that it is written the way you would say it to an actor. It can sit at the start of a line or partway through, and a sound effect such as a shaky breath is rendered where the bracket sits. A paired tag, such as an emphasis tag or a whisper tag wrapped around a few words, applies the direction to those words only and nothing else. In the web editor you select the words and pick a tag from a menu; in the API you write the tags into the text.

Why this matters for a marketer: voiceover revisions are usually about one phrase, not the whole read. "Slower on the price, lift the brand name" is a direction that a preset cannot express and a bracket can. The example is a 30-second product spot where the second sentence should drop to a confiding tone and return. With a preset you re-record the whole thing at a different setting and hope; with a bracket you mark the sentence.

The failure case is equally plain. A direction is an instruction to a model, not a guarantee, and Fish Audio publishes no figure for how reliably a given bracket produces a given delivery. A team that scripts a hundred spots with brackets and ships without listening has automated the mistake, not removed it. Listen to every take, or at least every take that carries a bracket.

The feature most likely to change a workflow is the one you cannot yet use. Audio inpainting, in Fish Audio's description, lets you mark a word in a finished take, type a replacement, and have the model regenerate only that word to fit the surrounding voice, mood and pace. For anyone who has re-recorded a two-minute explainer because a product name changed, that is the whole pitch. The launch post is explicit that it is in small, invite-only testing, alongside the relative-direction feature, and that an editing API is coming soon.

Read that as a roadmap item with a demo, not a shipped capability. It is also the kind of feature where the model's limits will show first: a replacement word with different syllable count or stress has to fit a slot the original take left, and the post gives no measure of how often the join is audible. When it reaches general availability it deserves its own test; until then, plan on regenerating the sentence.

The weights live in the fishaudio/drama-3 repository on Hugging Face, with an int4 quantisation alongside and the serving code on GitHub. Hugging Face's file metadata reads the main checkpoint as 36 billion parameters against the 35 billion in Fish Audio's description, which is the usual rounding gap between a marketing figure and a tensor count. The model page's own throughput numbers, measured by Fish Audio on a single NVIDIA H200, give the scale of hardware this expects.

Two readings of that table. First, a 3-billion-active-parameter model on a single data-centre GPU answers in under 100 milliseconds at low concurrency, which is the regime a real-time product needs; by 64 simultaneous streams the tail reaches a second. Second, every figure is the vendor's own run on the vendor's chosen hardware. Fish Audio publishes no VRAM minimum, and the int4 variant exists for a reason. A team planning to self-host should assume an H100-class card and measure, not a consumer GPU and hope.

The launch post summarises the licence in a sentence. We read the licence file, dated October 13, 2026 and headed Fish Audio Community License Agreement version 1.0, and the detail is where the decisions are. Research and non-commercial use are free. A commercial licence applies, in the licence's words, "only while Your Annual Revenue is less than USD 500,000", where annual revenue is the gross revenue of you and your affiliates from all sources over the trailing twelve months. Cross the line and you have 30 days to notify Fish Audio and 90 days of continued use before a separate agreement is required; falling back below it restores the community licence.

Three clauses matter more than the threshold. First, a hosted service needs a separate licence at any revenue: the licence defines it as making the model available so third parties can generate speech from their own text, audio or voice samples, through an API, SDK, hosting service, or an app whose primary purpose is voice generation. Second, the agency case is carved out in your favour: delivering only output you generated yourself, which the licence illustrates as an agency or studio producing voiceovers, dubbing or audiobooks for a client, does not require the client to meet the revenue threshold. Deploying the model itself for a client does, and the client's revenue is tested as if it were yours. Third, the use restrictions: no speech that imitates an identifiable person without consent, no impersonation to deceive or defraud, and no removing the model's watermark or provenance signal.

One small inconsistency to note for anyone auditing the repository: the Hugging Face page's metadata tag reads "fish-audio-research-license" while the page text and the licence file say Community License. The launch post says the older S2 models stay under the Research License, so the tag looks like a carry-over. Read the file, not the tag.

The failure case is the obvious one for agencies. A studio under the threshold builds a self-serve voiceover portal for its clients on the weights. The portal is a hosted service by the licence's definition and needs a separate agreement from day one, whatever the studio earns. The same studio delivering finished files is inside the licence.

Fish Audio's API price for Drama 3 is $50 per million characters at list. From October 13 to November 12, 2026 the launch post gives an 80% discount, which is $10 per million characters. On the creator side, Drama 3 draws on the same credits as S2.1 Pro, and new Plus, Pro or Max plans bought before November 12 are half price for the first month on monthly billing or the first year on annual. The post also offers a free month of the Plus plan for commenting on the launch within 48 hours, which is a promotion, not a price.

A third number comes from outside the vendor. Our October 15 capture of OpenRouter's catalogue lists a route named fish-audio/drama-3 at $2 per million UTF-8 bytes. The unit is bytes, not characters; for English text they are close, for accented and non-Latin scripts a character is two to four bytes. We could not read a price on OpenRouter's own model page when we fetched it, so the $2 figure stands on our catalogue capture alone, and the gap between it and the vendor's $50 list is not something we can explain from the pages read. Treat it as a number to confirm on the day you buy. Our guide to voice pricing units covers why characters, bytes, tokens and minutes do not convert cleanly.

A worked example at the vendor's list price. A 60-second spot runs about 150 words, or roughly 900 characters with spaces. A thousand such spots a month is about 900,000 characters, which is $45 at list and $9 during the launch window. The characters are our assumption and the prices are Fish Audio's; the point is that for short-form marketing audio the model cost is small next to the time spent listening to the takes, which is where section 02's warning lands.

The launch post's one quantitative claim is a pronunciation score. Fish Audio reports that on Artificial Analysis's English pronunciation robustness evaluation Drama 3 scores 93.2% overall against 88.1% for Fish Audio's previous model, S2.1 Pro, and that Drama 3 ranks first on that leaderboard. The post breaks the score into four categories: brands, places and technical terms; words whose pronunciation depends on context; shorthand such as numbers, dates and units; and exact sequences such as codes, paths and email addresses.

Two labels belong on that number. It is a third-party evaluation, which is better than a self-run benchmark. And it is cited by the vendor from the third party's leaderboard; we have not checked the leaderboard ourselves, and we do not know which other models Drama 3 ranked above on the day Fish Audio looked. The gain over S2.1 Pro is smallest on shorthand, under one point, and largest on brands and context-dependent words, where it is eight to nine points. For a marketer reading product names and prices aloud, that is the useful split. Our own pronunciation test for names and numbers is the method to run before trusting any of these figures on your own vocabulary.

The weakest category is the one that matters most for a technical product: exact sequences such as codes, file paths and email addresses, at 86.3%. If your scripts read out order numbers or URLs, that is the figure to test against, and a one-in-seven miss rate on a leaderboard is a one-in-seven retake rate in a studio.

No "best voice" verdict here; we have not run a listening test and the vendor's quality claims are the vendor's. What the licence, the price page and the feature list do support is a set of fits.

If your team produces campaign audio and is weighing a directable open-weight model against the hosted services it already pays for, the licence question and the listening test are both work our content engine team does, including the retake-rate measurement on your own product vocabulary.