Fish Audio Drama 3 Preview

Fish Audio Drama 3 Preview is an open-weight text-to-speech model developed by Fish Audio that specializes in multi-character scenes, long-form content, and expressive performance direction. Built around a mixture-of-experts (MoE) architecture, the model scales to 35 billion total parameters with approximately 3 billion parameters active per token. It is designed to handle complex voice narration by allowing users to direct performances using natural language instructions, inline emotional cues, and speaker tags.

Architecture and Capabilities

The model features a hybrid attention text backbone combined with a codebook-MTP audio decoder that decodes multiple codebooks causally over sequences. It supports up to 16 distinct speaker identities within a single script, making it suited for multi-character dialogue, audiobooks, and interactive media. In addition to multi-speaker capabilities, Drama 3 offers reference-based voice cloning from brief audio clips and supports streaming generation across 83 languages.

Prompt Guide and Control

Performance direction is achieved by inserting plain-language descriptions or tags within square brackets directly into the script, instructing the model on pacing, intent, and emotional nuance (such as whispering or shifting tone mid-sentence). The model interprets these contextual cues to alter delivery without requiring full speech regeneration, allowing fine-grained control over pronunciation and delivery.