S2 Pro

S2 Pro is Fish Audio’s text-to-speech model for generating expressive speech from text. Its dual autoregressive architecture combines a 4-billion-parameter slow stage for primary semantic codes with a 400-million-parameter fast stage for nine residual codebooks, balancing synthesis quality and inference speed. The report also describes GRPO-based post-training in which data-filtering models act as reward models.

The model is aimed at applications needing natural, controllable delivery rather than plain spoken output. Inline instructions can request effects such as whispering, laughter, pitch changes, or a broadcast style, with more than 15,000 tags covering emotion, tone, volume, and rhythm. It was trained on roughly 10 million hours of audio spanning more than 80 languages and does not require phoneme annotation or language-specific preprocessing. Reported results include low word error rates on Chinese and English Seed-TTS evaluations and strong multilingual speech evaluation performance, making it a practical fit for multilingual voice interfaces, narration, and expressive synthetic media.