Fish Audio S2.1 Pro (Oct 2026)

Fish Audio S2.1 Pro is a production-oriented text-to-speech and voice synthesis model developed by Fish Audio. Building upon the open-weight foundation of the S2 architecture, S2.1 Pro is designed for high-throughput, low-latency streaming applications and supports 83 languages. The model introduces substantial infrastructure optimizations, including custom CUDA kernels and continuous batching to maximize efficiency during autoregressive audio token generation.

Core Capabilities and Control

The model provides fine-grained inline emotion control and paralinguistic cues using flexible bracket syntax (such as [whisper], [laughing], or [emphasis]) embedded natively within text input rather than relying on a fixed set of classification tokens. S2.1 Pro also supports zero-shot voice cloning from short audio clips (typically around 10 to 15 seconds) and handles multi-speaker dialogue generation within a single continuous call, preserving speaker identity, timbre, and subtle micro-prosody across languages.