Fish Audio Drama 3 uses a 35B-parameter mixture-of-experts backbone with 3B active parameters per token to generate directed speech performances. The model supports multi-character scenes, natural-language direction, multi-turn revisions, and audio infilling via a hybrid linear-attention architecture.
HOW THIS AFFECTS YOU
●
builderYou can implement streaming audio and multi-turn voice editing using this MoE architecture.
●
designerYou can use natural language to direct specific emotional performances and character interactions in synthesized audio.