Fish Audio S2.1 Pro is a high-performance text-to-speech (TTS) and voice generation model developed by Fish Audio. Released in June 2026, it serves as a production-grade successor to the S2 Pro model, offering significant improvements in audio quality, latency, and system throughput. The model is built on a Dual-Autoregressive (Dual-AR) architecture, consisting of a 4B-parameter Slow AR backbone based on Qwen3 for semantic prediction and a 400M-parameter Fast AR for high-fidelity acoustic synthesis.
A defining feature of S2.1 Pro is its fine-grained inline control, which allows users to direct speech delivery using natural language instructions. By embedding tags in square brackets—such as [whispers], [laughing], or [professional tone]—users can trigger emotional or prosodic shifts at the word level without specialized SSML. The model supports 83 languages and is optimized for real-time performance, achieving a Time-to-First-Audio (TTFA) of approximately 70ms and a 2x throughput improvement over previous iterations.
The system also provides robust zero-shot voice cloning, enabling the creation of a digital voice clone from a 10-to-30-second reference clip. It utilizes a ModifiedDAC audio codec to ensure studio-quality 44.1 kHz output. In comparative evaluations, S2.1 Pro demonstrated a 61% win rate over S2 Pro, specifically excelling in the naturalness of multilingual synthesis and emotional expression.
For best results, users can leverage the model's ability to handle multi-speaker dialogues in a single generation. Natural language tags are not restricted to a fixed vocabulary, allowing for open-domain descriptions like [whispers sweetly] or [speaking with a slight cold] to influence the output. This flexibility makes it suitable for diverse applications ranging from immersive storytelling to low-latency conversational AI agents.