# Fish Audio Drama 3: A 35B Open-Weight Voice Model You Direct in Plain Language

- Published: 19 October 2026
- Updated: 19 October 2026
- Author: RenderBob team
- Category: Technical
- Canonical: https://renderbob.ai/blog/fish-audio-drama-3-open-weight-directable-tts

Fish Audio announced Drama 3 on 13 October. You write performance direction in square brackets, and one text can carry up to 16 speakers. The weights are on Hugging Face.

Fish Audio announced Drama 3 on 13 October. It is an open-weight text-to-speech model: you write performance direction in square brackets, and one text can carry up to 16 speakers.

## What Fish published

Fish's post and the technical report describe a 35 billion parameter mixture-of-experts model that activates about 3 billion parameters per token. Direction goes in plain language inside square brackets, anywhere in a line, and it can be in a different language from the spoken text. An emphasis tag applies the direction only to the words inside it. Speaker tags run from 0 to 15. Long text is chunked and can be streamed. Weights are on Hugging Face, with hosted apps and an API as the alternative.

A ComfyUI Wiki write-up puts the download at about 40GB, lists 83 languages, and says the partner nodes call Fish Audio's API and need an API key. The Fish post we read gives no disk size, no language count, and no GPU requirement.

## The licence

The Hugging Face licence is the Fish Audio Community License v1.0, dated 13 October. Research and non-commercial use are free. Commercial use is free while you and your affiliates stay under $500,000 in annual revenue. Offering Drama 3 to other people as a hosted speech service, a voice-cloning service, or an inference API needs a separate licence from Fish Audio. The licence also bans cloning a real person's voice without their consent.

## The pronunciation score

Fish says Drama 3 ranks first on Artificial Analysis' Pronunciation Robustness benchmark, at 93.2% against 88.1% for its previous S2.1 Pro model. Fish calls the test independent. The number is still Fish citing it. On names of brands, places, and technical terms, Fish reports 95.9% against 87.3%.

## Beside Eleven v4

Our 13 October piece on Eleven v4 described inline audio tags such as [laughs] and [whispers]. Both products let you direct the read in the text. Drama 3 ships weights you can host. v4 is a hosted service, and ElevenLabs says every cloned voice needs verified owner consent.

## Consent

The technical report says the model clones voices from references. The licence forbids doing that with a real person's voice unless they consent. Get the permission in writing, including the scope of use, and keep it with the project. That is the same record our ElevenLabs note asks for.

## A quick test

Write one scene with two characters. Direct a mid-line change of tone on one of them. Generate it three times and compare the voice and the pronunciation, especially on names. When that voice goes into a picture cut, RenderBob lets you choose local or cloud for each clip and keep the references with the studio.

## What to watch

How productions read the $500,000 revenue line, a published hardware requirement for local use, and listening tests of direction control across languages.