Fish Audio S2

Fish Audio S2 is a text-to-speech system released on March 10, 2026, positioned as the flagship of the Fish Speech ecosystem. It uses a hierarchical Dual-Auto-Regressive design with a roughly 4-billion-parameter slow stage for linguistic planning and a 400-million-parameter fast stage for acoustic rendering, combined with residual vector quantization of 44.1 kHz audio.

The headline feature is instruction-following emotion control: users direct delivery with natural-language tags for laughter, whispers, or moods, alongside zero-shot voice cloning from short samples. Fish Audio releases the model weights, fine-tuning code, and an SGLang-based streaming inference engine, with code and checkpoints on GitHub and Hugging Face. Verify the exact weight license on the model card before commercial use.