The company Fish Audio published a preview of the new speech model Drama 3, replacing short audio tags with descriptive prompts.

Primary change in prompting

Instead of strict one‑ or two‑word audio tags, creators can now describe intonation, tempo and character in ordinary text prompts to shape performance.

Key capabilities

The preview highlights several functional improvements compared with earlier releases and demonstrates more flexible control over generated audio.

• Selective regeneration: the engine can re‑render a single word or phrase inside an existing track without altering the surrounding audio.
• Mid‑utterance voice switching: the model is able to change speaker timbre within a single sentence or phrase.
• Multi‑speaker dialogue: Drama 3 can produce conversations of several characters in one pass.
• Russian language support: the system generates speech in Russian alongside other supported languages.

Comparison with version 2.1

Testing with version 2.1 revealed that some tags did not always affect output, and stress placement occasionally missed the mark during synthesis.

In contrast, using the same prompt with a few descriptive phrases in Drama 3 produced clearer enunciation and more natural intonation throughout sentences.

Some crowd‑reaction cues such as applause or cheering were still ignored in the preview, but overall delivery showed noticeable refinement over the previous model.

Availability and attribution

According to communications from Fish Audio, the preview build of Drama 3 is already accessible for free on the creator’s site, with broader release plans to follow.

During earlier tests, available voices included a speaker identified as Sergey Burunov, whose sampled timbre resembled the actor’s performance in those trials.