Fish Audio S2 is an open-source TTS (Text-to-Speech) model released by Fish Audio in March 2026. In a nutshell, it's a multilingual voice generator that allows you to inline specify laughter, whispering, and professional broadcast tones using 15,000 emotion tags. Existing zero-shot TTS models either read text in a monotonous tone or required a separate prosody model to add emotion. Fish Audio S2 uses a Dual-AR architecture (4B Slow AR + 400M Fast AR) to separately predict the semantic codebook and residual voice codebook, allowing it to receive emotion and prosody tags in a natural language inline format. You can incorporate tone changes within a single sentence, such as [laugh] Hello [professional broadcast tone] This is today's news, and it achieved a Bradley-Terry score of 1st place (3.07) in a blind evaluation. From a practical perspective, it (1) covers 80+ languages with a single model, allowing you to handle multilingual content dubbing with one model; (2) creates zero-shot voice clones with only 10-30 seconds of reference audio, generating consistent speaker voices for each character; and (3) with an RTF of 0.195 (H200 1 card) and TTFA of ~100ms, it can be directly integrated into the backend of interactive voice agents for real-time inference.

Dual-AR architecture, 4B parameter, high-quality text-to-speech model.

Real-time Multilingual Text-to-Speech: Process over 80 languages with a single model, enabling global content localization. Clone native speakers with a 10-second reference audio and automatically generate multilingual dubbing. With an RTF of 0.195, it can be deployed in real-time for streaming services.