Fish Audio releases the Drama 3 speech model: controlling performance with parenthetical stage directions
Fish Audio Drama 3: Directable AI Voice Acting Explained
Fish Audio released the text-to-speech model Drama 3, allowing users to write stage directions in square brackets within scripts to adjust tone, pacing, and performance. It also supports word-level control, inline sound effects, and multi-speaker scripts. The model is released as open weights with 36B parameters on Hugging Face (fishaudio/drama-3). According to user reports, full-precision operation requires approximately 46GB of VRAM, while the community 4-bit quantized version requires approximately 32GB.
Fish Audio open-sources a 36B speech model, using bracketed stage directions to control performance, and provides benchmark differences compared to ElevenLabs along with self-hosting VRAM requirements.
Open original mindstudio.ai