Fish Audio releases S2 Pro, a 5B TTS model for 80+ languages, research licence
S2 Pro is a dual-autoregressive TTS model with inline emotion tags; weights are free for research, commercial use needs a licence.
Fish Audio released S2 Pro, a text-to-speech model with a roughly 4B slow autoregressive stage and a 400M fast stage, listed at 5B parameters. The card says it supports 80+ languages, with Japanese, English and Chinese in the top tier. Inline tags such as [tag] take free-form descriptions of prosody and emotion, and the card claims over 15,000 unique tags. It says training used over 10 million hours of audio with reinforcement learning alignment.
The release includes weights, fine-tuning code and an SGLang-based streaming inference engine. The card has no quality benchmarks. Its only figures are speed on one NVIDIA H200: a real-time factor of 0.195 and about 100 ms to first audio.
The licence is the Fish Audio Research License. Research and non-commercial use is free; commercial use requires a separate licence from Fish Audio.
The card gives no release date. The date used is the Hugging Face repo creation date, 9 March 2026; the card links a technical report published on 9 March.
Plenty of speed numbers. No quality numbers. Others' tables will have to do.
Sources
Editor's first pass (AI-assisted)
- Real 4: Weights, fine-tuning code and a streaming engine are released.
- Useful 3.5: Free-form tags for emotion help; commercial use needs a separate deal.
- Notable 3.5: Fish Audio is a known TTS name; higher scores on rivals' cards than on its own.
- Novel 3: Dual-AR with free-form tags is a refinement of its earlier line.
Comments