What is Fish Audio S2?

Fish Audio S2 is a next-generation text-to-speech model that brings unprecedented expressiveness to voice AI. Unlike traditional TTS systems that produce flat, robotic speech, S2 lets you control emotion, tone, and delivery using natural language instructions embedded directly in your text. The model is fully open-source, including both inference code and model weights, making it accessible for developers, researchers, and creators who want to build realistic voice applications without vendor lock-in.

Who it's for

• Content creators – who need to generate natural-sounding voiceovers with emotional nuance, laughter, or whispers for videos, podcasts, or audiobooks
• AI developers – building conversational agents, virtual assistants, or interactive voice applications that require sub-150ms latency and real-time responsiveness
• Multilingual teams – working across 80+ languages who want consistent, high-quality speech generation with fine-grained control over speaker identity and delivery

Key features

Ultra-low latency for real-time use

Fish Audio S2 delivers speech generation in under 150ms, enabling seamless conversational AI, live dubbing, and interactive voice experiences. The SGLang-based inference engine supports continuous batching and prefix caching, making it production-ready without sacrificing quality.

Open-domain control with natural language tags

You can direct the voice by adding simple tags like [whisper], [laughing nervously], or [professional broadcast tone] directly in your text. Over 15,000 unique tags are supported, giving you word-level control over emotion, emphasis, pitch, and paralanguage without needing complex parameters.

Multi-speaker dialogue in a single pass

Switch between speakers naturally within one generation using the <|speaker:1|> syntax. This makes it easy to create realistic conversations, dramatic readings, or multi-character audio without stitching separate clips together.

Fully open-source architecture

Both the 4B-parameter semantic model and 400M-parameter acoustic model are released under the Fish Audio Research License. You can run S2 on your own hardware, fine-tune it on custom data, and integrate it without API dependencies or recurring costs.

What stands out

"The most expressive voice AI ever made, now open-source."

Fish Audio S2 redefines what's possible with text-to-speech by treating voice direction as a natural language problem. Instead of choosing from a handful of preset emotions, you can describe exactly how you want the voice to sound — from a barely audible whisper to an excited shout — and the model interprets it correctly. Combined with multi-speaker support and 80+ language coverage, this makes S2 a genuine platform for building lifelike voice experiences, not just another TTS API.

Worth checking out if…

You're building any application where voice quality and emotional authenticity matter — whether it's a conversational AI agent, a multilingual dubbing pipeline, or an interactive storytelling tool. Fish Audio S2 is especially valuable if you want full control over your voice infrastructure without being locked into a proprietary service.