Fish Audio Drama 3 lets AI voices change emotion mid-sentence

Fish Audio has launched Drama 3, a text-to-speech model designed to make AI-generated dialogue sound less like someone reading a script and more like an actual performance. Released on October 13th, the model lets creators control emotions, pacing, and vocal delivery using ordinary language.

Instead of relying on basic voice presets, users can insert instructions directly into their scripts. A character might begin speaking calmly, become increasingly frustrated, and finish a sentence in a whisper, all without switching between separate recordings. Even synthetic actors now have to take direction.

Drama 3 also supports multi-character conversations and 83 languages. Instructions can be applied to individual words or passages, allowing creators to emphasize specific expressions without changing the rest of the script.

According to Fish Audio, the model achieved 93.2% accuracy on Artificial Analysis' Pronunciation Robustness benchmark, compared with 88.1% for its predecessor, S2.1 Pro. However, pronunciation accuracy does not necessarily guarantee convincing emotional performances.

The company is also developing two advanced editing tools. Interactive Editing will allow users to refine performances through follow-up instructions, while Audio Inpainting is designed to replace individual words without regenerating entire recordings. Both remain in limited, invitation-only testing.

Drama 3 is available through Fish Audio's website, mobile apps, and API. Model weights are also downloadable, although commercial use is subject to licensing restrictions.

Until November 12, API access costs $10 per million characters, an 80% discount from the regular $50 rate. New creator subscriptions are also discounted by 50% during the promotional period.