How to Run Fish Audio S2 Pro Locally: Multilingual TTS with Inline Emotion Control

Fish Audio S2 Pro is a 4.56 billion parameter open-weight text-to-speech model released March 9, 2026 by Fish Audio. It speaks 80+ languages and gives you fine-grained control over prosody and emotion by embedding natural-language tags directly in the text you want synthesized. The model ships with weights, fine-tuning code, and an SGLang-based streaming inference engine.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts and hardware requirements against the primary source before recommending an install path.

A top-20 entrant on Hugging Face's trending list

S2 Pro entered the top 20 of Hugging Face's text-to-speech trending list. At time of writing it has 1,379 likes and 47,024 downloads on the fishaudio/s2-pro model page, plus 32,863 stars on the fishaudio/fish-speech GitHub repo. There is also a community demo Space, artificialguybr/fish-s2-pro-zero (184 likes), if you want to hear a sample before pulling the weights.

Specs at a glance

Hardware: what 4.56B parameters actually costs you

Fish Audio's own install docs list 24GB of GPU memory as the requirement for inference, and that is the number to plan around rather than a bf16 weight estimate. The weights alone total roughly 9.1GB at bfloat16, but the streaming pipeline, KV cache, and the 400M fast-AR pass at every timestep add enough overhead that Fish Audio does not recommend going below 24GB. That puts a 3090, 4090, or an A5000-class card in the realistic range for local use. CPU inference is possible in principle but the 4B slow-AR component makes it impractically slow for anything beyond a test clip.

The production benchmark on a single H200 shows a real-time factor of 0.195, time-to-first-audio under 100ms, and throughput above 3,000 acoustic tokens per second. Those numbers will not transfer to a consumer card, but they give you a sense of how much headroom the architecture has when hardware is not the constraint.

Architecture: why it is faster than it looks

S2 Pro uses a Dual-AR design split into two components that run at different speeds. The slow AR is a 4B-parameter transformer that works along the time axis and predicts the primary semantic codebook. The fast AR is a 400M-parameter model that fills in the remaining 9 residual codebooks at each time step, recovering fine acoustic detail without running the full 4B model for every token.

Because this architecture is structurally isomorphic to a standard autoregressive LLM, it inherits all the serving optimizations that SGLang supports out of the box: continuous batching, paged KV cache, CUDA graph replay, and RadixAttention prefix caching. That is where the low RTF comes from.

Inline emotion and prosody control

S2 Pro does not use a fixed list of speaker IDs or a set of predefined style presets. Instead you embed natural-language instructions inside the text using bracket tags. The model was trained on 15,000+ unique tags, so you are not limited to a closed vocabulary.

Common examples from the model card: [whisper in small voice], [professional broadcast tone], [pitch up], [pause], [emphasis], [laughing], [excited], [angry], [sad], [interrupting]. You can place them at the word level, not just at the start of a sentence, which lets you change the delivery mid-utterance rather than applying a single style to the whole clip.

Install it

Clone the fish-speech repo, install the CUDA dependencies, and bring up the web UI or use the CLI. The project uses a Docker Compose file for the full stack.

FAQ

What is Fish Audio S2 Pro?

A 4.56 billion parameter open-weight text-to-speech model from Fish Audio that supports 80+ languages and lets you control prosody and emotion at the word level using plain-text tags embedded in your input.

How many parameters does it have?

4.56 billion, per the safetensors metadata in the Hugging Face API. The model splits into a 4B slow-AR component and a 400M fast-AR component.

What GPU do I need?

Fish Audio's install docs state 24GB of GPU memory for inference. That covers a 3090, 4090, or similar. Cards below 24GB are not the supported configuration and are likely to hit memory limits during streaming.

Which languages does it support?

80+ languages. Tier 1 (best quality) is Japanese, English, and Chinese. Tier 2 includes Korean, Spanish, Portuguese, Arabic, Russian, French, and German. Dozens of additional languages are supported at varying quality levels.

Can I use it commercially?

Not for free. The Fish Audio Research License permits research and non-commercial use at no cost. Commercial use requires a separate license from Fish Audio; contact business@fish.audio.

How fast is it on consumer hardware?

Fish Audio's published benchmark on an H200 shows an RTF of 0.195 and time-to-first-audio under 100ms. Consumer GPUs will be slower, but the Dual-AR architecture and SGLang backend are designed to keep latency low relative to model size.

Where to go from here

• Need a fully open TTS model for commercial work? Run Kokoro-82M locally covers a smaller, Apache-licensed alternative.
• Want Indian-language voice cloning from a tiny checkpoint? Run IndicF5 locally covers a 350M-parameter model that handles 11 Indian languages.
• Choosing a GPU for a local-AI box? Local AI on 24GB of VRAM covers the hardware tier S2 Pro's own docs call for.
• Compare S2 Pro against other open TTS models: browse the local AI model directory to see specs side by side.