Fish Audio unveils Drama 3: AI voices that take stage directions
The open voice model covers 83 languages and follows directions written in square brackets. Smaller companies may use it commercially for free.
Graphic: TERRA WORLD NEWS (own graphic based on the sources listed)
Video: YouTube · Fish Audio · youtube.com
AI start-up Fish Audio released its new text-to-speech model Drama 3 on Tuesday. Users write stage directions in square brackets right next to the text, and individual words can additionally be tagged, for example for emphasis. The model speaks 83 languages, and directions can be written in any language. Chief executive Rissa Cao calls Drama 3 the company's most emotive and controllable model yet.
Technically, Drama 3 is a mixture-of-experts model with 35 billion parameters, of which about three billion are active per token; it is built on Alibaba's Qwen3.5-35B-A3B language model. On a pronunciation test by Artificial Analysis it scores 93.2 per cent according to the company, compared with 88.1 per cent for its predecessor S2.1 Pro. On a single Nvidia H200 chip, the model generates 271 audio frames per second, Fish Audio says – considerably more than S2.
The weights are available on Hugging Face under the company's own community licence: free for research and non-commercial use and for companies with less than 500,000 dollars in annual revenue. Larger companies, and anyone offering Drama 3 as a speech, voice-cloning or inference service, need a licence. Real people's voices may not be cloned without their consent; according to Fish Audio, the launch video features a cloned voice of comedian "Uncle Roger" – used with his permission. Drama 3 is also available via the website, apps and API.