Fish Audio releases Drama 3, TTS model with acting cues
Direct an AI voice actor in text
You can now direct a voice actor by typing. Write [barely holding it together] before a line, and Fish Audio's new Drama 3 reads it that way.
Drama 3 is a text-to-speech model that turns a script with bracketed acting cues into spoken dialogue. Paired tags such as <whisper> change only the words they enclose. The directions can be in any language, whatever language the line is spoken in, and the model supports 83 languages.
Up to 16 characters can share one request. The API takes one voice ID per speaker, so a full scene renders in a single call. The weights are on Hugging Face and the model is live in the Fish Audio API.
On pronunciation, Fish Audio reports 93.2% on Artificial Analysis Pronunciation Robustness, ranked #1, up from 88.1% for S2.1 Pro. On the quality leaderboard, Drama 3 Preview scores 1242 Elo, behind Eleven v4 Turbo at 1328 and ahead of S2.1 Pro at 1190.
The weights are free for research and non-commercial use, and free commercially under $500,000 in annual revenue. Self-hosting needs an NVIDIA GPU with at least 48 GB of memory in FP8. The hosted API lists at $50 per 1M characters, with a $10 launch price through November 12, 2026.
Would you write dialogue this way?