Fish Audio open-sources Drama 3, a 35B MoE speech model decoding at 271 fps on one H200
fishaudio · hf · 2026-10-14
Fish Audio released Drama 3, an open-source speech generation model on a 35B MoE backbone (3B active per token) that turns text into directed performances: multi-reference voice cloning, natural-language direction, take revision from notes, multi-character scene dubbing, infilling-based speech editing, plus long-form and streaming synthesis — the latter four new in Drama 3. It pairs a hybrid linear-attention backbone with a codebook multi-token-prediction audio decoder, trained via three-stage pretraining, instruction tuning and large-batch RL with a single learned human-preference reward. It decodes one stream at 271 fps on one H200 vs 177 for the 7x-smaller Fish Audio S2. Weights are on GitHub and Hugging Face.
More from Multimodal
- AI-generated "boring festival" series with consistent character, full Magnific One prompt shared — techhalla · 2026-10-17
- Boring set saved by two standout AI image generations, prompts shared — techhalla · 2026-10-17
- Reddit user unimpressed after trying Seedance 2.5 video model — Due-Speaker5476 · 2026-10-17
- MOSS easiest model for generating new faces per seed, says Reddit user — JohnToFire · 2026-10-17
- AI video creator cuts reroll costs 87%: draft at 480p, upscale keeper to 4K — thetripathi58 · 2026-10-17
- How to get consistent character voices in AI video? A Minimax workflow question — brittpitre · 2026-10-17