Fish Audio releases Drama 3 with script-directed voices and downloadable weights

A scripted AI voice can take new direction halfway through a line, shifting from restrained anger to a whisper without changing the words. That is Fish Audio’s pitch for Drama 3, released on October 13, 2026. The speech model is available through the company’s website, mobile app and API, with downloadable weights for teams that want to run it themselves.

The script carries the performance

Drama 3 takes ordinary-language instructions in square brackets alongside dialogue. A direction can establish the mood at the beginning, change it later, or place a breath or laugh at a particular point. Paired tags offer narrower control: wrapping a few words in a whisper tag applies that instruction only to those words, rather than everything that follows.

The notation is not itself the new feature. Fish Audio says users can give Drama 3 the same directions as its earlier S2.1 Pro model, but get a more expressive performance and closer instruction-following. Its examples emphasize emotions that can be flattened in generated speech: crying while trying to smile, holding back anger, and breaking into laughter.

Fish Audio says the model speaks 83 languages, and the language of a direction need not match the dialogue. Its launch examples apply Chinese, Spanish and Japanese instructions to an English line. The company positions Drama 3 primarily for recordings that creators generate and keep, including audiobooks, audio drama, game dialogue, animation and short videos.

Generating a performance and repairing an approved recording are separate capabilities at this launch. Interactive editing and audio inpainting remain in small, invite-only tests, with an editing API planned. Inpainting is designed to replace a selected word while matching the voice, mood and pace of the surrounding audio, leaving the rest of the take intact.

Expression is not the pronunciation test

Fish Audio reports that Drama 3 scored 93.2% on Artificial Analysis’s Pronunciation Robustness benchmark, compared with 88.1% for S2.1 Pro. The test puts difficult words, numbers, codes and context-dependent readings into English sentences, then uses human listeners to check them. That measures whether the words come out correctly, not whether a dramatic scene achieves its intended emotional effect.

Fish Audio reports a larger improvement on names and technical terms than on shorthand such as numbers and dates. Exact sequences improved in the company’s published results, but remained less accurate than the other listed categories.

The company describes a training change aimed at preserving expression. Earlier speech models, including its S2, used rewards for easy transcription and similarity to a reference recording; those scores could penalize a requested laugh or whisper. Drama 3 instead uses one learned reward model based on which takes people preferred. That evaluator also receives the script and voice, so accuracy and likeness still count.

Underneath, Drama 3 has 35 billion parameters, the model’s learned settings, but activates about 3 billion for each generated token. This mixture-of-experts design uses a subset of its components at each step. Fish Audio also describes a new audio decoder that draws on the preceding sequence, including reference audio, when generating the fine detail of sound.

Downloadable does not mean unrestricted

The weights are available on Hugging Face for self-hosting under the Fish Audio Community License v1.0. Commercial use is free only while a business and its affiliates have less than $500,000 in annual revenue. The license also draws a separate line around offering the model to other users as a service.

• Research and non-commercial use are free under the community license.
• Businesses above the revenue threshold need a license from Fish Audio. So do providers offering Drama 3 as a speech-generation, voice-cloning or inference service or API.
• Cloning a real person’s voice without consent is prohibited.

For hosted access, Fish Audio’s launch offer lists the API at $10 per million characters from October 13 through November 12, discounted from a $50 list price. Its current instructions use the model header drama-3 at the same speech endpoint as S2.1 Pro. The company says an existing integration can switch models by changing that header.

The consumer version uses the same credits as S2.1 Pro. Fish Audio also acknowledges that takes vary and recommends generating several before choosing one. That leaves listening and selection inside the workflow, even when the script gives precise directions.