Drama 3 Turns Dubbing Into Stage Direction

Fish Audio released Drama 3 on October 13, and the first thing most people noticed was the size: 35 billion parameters, mixture of experts, about 3 billion active per token. The second thing was the benchmark number, 93.2% on pronunciation robustness, up from 88.1% on the previous S2.1 Pro, and reportedly enough to put an open-weights model at the top of that particular table.
Both of those are worth a paragraph and no more. The thing that actually changes how voice work gets done is the input format.
You no longer type a sentence. You write a scene note.
Old text-to-speech had one job: take characters, return audio. Emotion was something you chased by picking a different voice, nudging speed, and rolling the dice until a take sounded right.
Drama 3 takes directions in square brackets, in plain language, inline with the script. Fish's own example puts four of them into a single line: hold it together and say it is fine, then a shaky breath, then a small laugh, then drop into a whisper. The words never change. Everything that makes the line actable lives in the brackets.
There is a second layer. Paired tags wrap individual words, so a passage can stay flat and one phrase inside it can go breathy. Sighs, laughs, and pauses go where they belong instead of being pasted in during post.

If you have ever directed a human actor, this will look familiar. It is a script with margin notes. If you have only used TTS, it is a strange way to work for about ten minutes and then it is the only way you want to work.
Multi-speaker in one pass
The other structural change is that a single generation can hold up to sixteen speakers. Not sixteen clips stitched together — one run, with voices that stay distinct and, according to Fish, respond to each other's tone.
Anyone who has assembled a dialogue scene from separate TTS clips knows what this removes. You no longer spend an afternoon matching room tone between line 4 and line 5, or re-recording a whole exchange because one voice drifted. Long-form consistency is the same story: the voice and the pacing hold across a script instead of resetting every few thousand characters.
83 languages, one model, no per-language endpoint. Switching language inside a scene is supported, which is the part that matters for anyone localizing the same piece of content into four markets.
What it does to the cost of changing your mind
Here is the quiet economic argument. In a traditional recording session, fixing one word means re-recording the line, and re-recording the line means the timing, the breath, and the emotional arc all shift slightly. So you either accept a seam or you redo the scene.
Drama 3 ships an audio inpainting feature that regenerates a selected segment without touching the performance around it. In theory, that collapses the cost of a one-word fix to near zero.
In practice, at launch, it is gated. Reporting on the release says single-word inpainting was in invitation-based testing, with a public editing API only planned. That is the most valuable feature in the announcement and you probably cannot use it yet. Fish also acknowledges that the same direction can produce different results across takes, and recommends generating several and picking one — which tells you this is still a tool that needs a human ear, not a one-click button.
The three things that will stop you
Hardware. In FP8 you need roughly 48GB of VRAM on an NVIDIA card. No CPU, no Apple MPS, no AMD. Someone in the community squeezed it into 4-bit on a Mac within a day, but that is tinkering, not production. For most teams the realistic path is the API or a platform integration.
License. The weights are on Hugging Face under the Fish Audio Community License v1.0. Free below $500,000 in annual revenue. Above that, you negotiate. If you want to resell it as a hosted voice service, that is a separate agreement regardless of your revenue.
Benchmarks you cannot check. The 93.2% figure and the "number one among open weights" placement come to most readers second-hand. The leaderboard page, the methodology, and the model card are not all in the same place, and ElevenLabs has not responded. Pronunciation robustness is a real and measurable thing — names, numbers, and foreign words read correctly on the first attempt — so treat the direction as credible and the exact ranking as provisional.
How to actually try it
Fastest route is the hosted API. Send text plus a reference_id for a cloned voice, pick a format and bitrate, get audio back. List pricing sits around $15 per million characters, with a promotional rate running for part of October and November. A 30-minute podcast lands somewhere near forty cents at that rate, which is the kind of number that makes you stop treating voice as a fixed cost.
If you want the weights, start with the model card, check your VRAM against the FP8 requirement, and read the license before you build a product on it. The 35B size is not a marketing figure; it is the reason this needs a rented card rather than a laptop.
Test it in this order
If you have an hour, spend it on the failure cases rather than the showcase lines.
Start with a sentence containing a name, a number, and an acronym in the same clause — the exact shape that makes TTS engines stumble. Then add a bracket direction and see whether the correction survives the difficult word. If the name still comes out right, the pronunciation claim is doing something real.
Next, run the same line five times and listen to the variance. Fish says results differ between takes and suggests generating several. That is honest, and it also means your workflow needs a selection step. If the five takes are wildly different, the direction is a suggestion rather than a control.
Then try a two-speaker exchange where one character interrupts the other. Multi-speaker generation is easy to demo and hard to land, because interruption requires the model to model overlap rather than take turns. This is where most multi-voice systems sound like a table read instead of a scene.
Finally, switch language mid-scene. Consistency across a language change is the feature that matters if you are localizing, and it is the one nobody posts examples of.
Write down what breaks. The list is more useful than any benchmark percentage, because it tells you which jobs to route to this model and which to keep in a booth.
Why this one matters more than the score
Voice generation has been good enough for narration for about two years. What it has not been good enough for is performance — the difference between reading a line and meaning it. That gap was closed by human directors in a booth, and by editors who cut around the flat takes.
Drama 3 does not close that gap. It moves the controls into the text, which means the person writing the script is now also the person directing the read. For a podcast, a game, an audiobook, or a dubbed short, that is a real shift in who does the work and when.
The tool is not finished. The best feature is behind a waitlist, the hardware bar is high, and the benchmark needs independent confirmation. But the interface is right, and interfaces are the part that tends to stick.
Related articles
AMD Got 19 Percent for Being the Other Option
Money did not leave AI. It left the one name with a governance problem and went to the two without one.
The $99 Billion Loop Nvidia Built Around Its Own Customers
Selling shovels is a good business. Lending money to the people buying shovels is a different one.
The 74 Percent Gross Margin Has a Shelf Life
Three forces arrive in the same four quarters, and each one costs about $3 billion per point.
175,000 Signatures Will Not Put Him Back. Here Is What Would.
The longer the campaign takes to win, the weaker its case becomes. That is the trap built into it.