Fish Audio Drama 3

Fish Audio Drama 3 is an open-source text-to-speech and voice generation model designed for multi-character scenes, directional control, and long-form narrative content. Built on a 35B-parameter mixture-of-experts (MoE) text backbone that activates 3B parameters per token, the model allows creators to guide performance dynamics using natural-language instructions and bracketed annotations directly inside the script. It supports multi-speaker generation, zero-shot voice cloning, and precise word-level tags for sounds, pauses, and emotional delivery.

Architecture and Capabilities

The model utilizes a hybrid framework incorporating a large-scale MoE text encoder coupled with an advanced codebook decoder that produces multiple codebooks per frame, eventually decoded into high-fidelity 44.1 kHz mono audio. Drama 3 is engineered to manage complex multi-turn dialogue scenes in a single pass, accommodating up to 16 distinct character voices while preserving emotional continuity and character tone across extended audio scripts.

Prompt Guide and Usage Tips

To direct the model's performance, users can insert plain-text descriptive tags within square brackets alongside the dialogue script—such as [whispering] or [voice catching, trying to stay steady]—to control the emotional nuance and pacing of specific words or lines. For best results in long-form generation, scripts should clearly delineate speaker tags and leverage inline directions to handle pauses and sound cues dynamically.