S2 Pro
Fish Audio's 4.6B text-to-speech model with free-form inline tags and voice cloning from 10 to 30 seconds. Research and non-commercial only.
In short. S2 Pro reads text aloud with acting directions you write in square brackets, like [whisper] or [laughing], and it can copy a voice from a 10 to 30 second clip. It needs a 24 GB GPU, and the license covers research and non-commercial use only unless Fish Audio gives you a commercial license.
Summary
Fish Audio released S2 Pro as open weights on 9 March 2026, and its GitHub page now calls it the previous-generation flagship, after Drama 3 in October. It is a dual-autoregressive text-to-speech model. A 4B "slow" transformer predicts the main semantic codebook along the time axis, and a 400M "fast" transformer fills in the other nine residual codebooks of an RVQ audio codec at each step, which comes to 4.56B parameters in BF16 plus a 1.9 GB codec. Instead of a fixed tag list it takes free-form instructions in square brackets, such as [whisper in small voice] or [professional broadcast tone], and the card says more than 15,000 distinct tags are supported. It also does multi-speaker dialogue with speaker IDs and cloning from 10 to 30 seconds of reference audio. The card says it was trained on more than 10 million hours of audio in 80+ languages, with Japanese, English and Chinese as tier 1.
Fish Audio reports a Seed-TTS Eval word error rate of 0.54% for Chinese and 0.99% for English, which it says is the best among the systems it tested, including Qwen3-TTS at 0.77 and 1.24. It also reports an Audio Turing Test score of 0.515 against 0.417 for Seed-TTS, an 81.88% win rate on EmergentTTS-Eval against a gpt-4o-mini-tts baseline, and the best word error rate in 11 of 24 languages and the best speaker similarity in 17 of 24 on the MiniMax multilingual test set. All of these are Fish Audio's own results. On one NVIDIA H200 the card gives a real-time factor of 0.195, about 100 ms to the first audio and more than 3,000 acoustic tokens per second while staying under a real-time factor of 0.5.
Fish Audio's install guide asks for at least 24 GB of GPU memory, on an NVIDIA GPU or a CPU, under Linux or WSL, and the project points to SGLang-Omni and vLLM-Omni for serving. The files total about 11 GB. The license is the Fish Audio Research License Agreement, which allows research and non-commercial use and requires a separate written license for anything commercial, including internal business use. Drama 3 uses the more open Fish Audio Community License instead. In the 30 days to 11 November 2026, Hugging Face counted about 47,000 downloads, and the repository has about 1,500 likes.
Good for
- Research, hobby projects and evaluation of expressive text-to-speech with inline emotion and delivery tags
- Prototyping multi-speaker dialogue and audio drama, with speaker IDs and cloned voices from 10 to 30 second clips
- Speech in Japanese, English, Chinese and many other languages, after you test the ones you need
- A single 24 GB NVIDIA GPU (the guide also lists CPU) rather than the 48 GB that Drama 3 asks for
Watch out for
- Commercial use is not allowed without a separate written license from Fish Audio, and that includes your own company's internal use, a paid product and a hosted service or API
- The weights and outputs may not be used to train other foundation models, and Fish Audio's Acceptable Use Policy applies
- Fish Audio gives three language counts, about 50 in the blog introduction, 80+ on the card and about a hundred in the GitHub README, so test your language before you commit
- All benchmark numbers are Fish Audio's own, and Fish Audio says Drama 3 adds local editing and multi-turn revision that S2 Pro does not support
- The card says nothing about consent for cloned voices or about watermarks, so labeling and consent are up to you
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of S2 Pro. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| S2 Pro BF16 | 4.56B | 11 GB | 24 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 24 GB of GPU memory.
The license in plain words
Fish Audio's license for the S2 Pro weights, last updated on 7 March 2026. Research and non-commercial use is free, including hobby projects, evaluation and testing. Any commercial purpose, including your own organization's internal operations, needs a separate written license from Fish Audio. If you share the weights or a derivative, you pass on the agreement, keep a notice crediting 39 AI, Inc. and show "Built with Fish Audio".
- Attribution or model name must be shown
- Acceptable-use policy applies
More about Fish Audio Research
See also
- S2 Pro vs Higgs TTS 3: which non-commercial TTS model?
- Best open-source text-to-speech models in 2026
Frequently asked questions
- What is S2 Pro?
- S2 Pro reads text aloud with acting directions you write in square brackets, like [whisper] or [laughing], and it can copy a voice from a 10 to 30 second clip. It needs a 24 GB GPU, and the license covers research and non-commercial use only unless Fish Audio gives you a commercial license.
- Can I use S2 Pro commercially?
- Not without separate permission. S2 Pro is released under the Fish Audio Research, which allows research and personal use only.
- Is S2 Pro open source?
- S2 Pro is open-weight, not open source in the strict sense. You can download the model files, but the Fish Audio Research adds conditions that the Open Source Initiative definition does not allow.
- Can I run S2 Pro on my own computer?
- Yes, on a good graphics card. The smallest variant, S2 Pro, needs about 24 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was S2 Pro released?
- S2 Pro was released by Fish Audio on March 9, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Fish Audio S2 Pro model card (Hugging Face) huggingface.co
- Fish Audio Open-Sources S2: Fine-Grained Control Meets Production Streaming (Fish Audio blog, 9 Mar 2026) fish.audio
- Fish Audio S2 Technical Report (arXiv 2603.08823) arxiv.org
- Fish Audio Research License Agreement (Hugging Face) huggingface.co
- Fish Speech installation requirements speech.fish.audio
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Drama 3 Fish Audio | 35.9B 3B active | – | Fish Audio Community 1.0 (Conditional) | 25 GB Gaming GPU | |
| Qwen3-TTS Alibaba Qwen | 910M – 1.92B | – | Apache 2.0 (Permissive) | 3.68 GB Laptop or phone | |
| Higgs TTS 3 Boson AI | 4.65B | – | Boson Higgs Non-Commercial (Non-commercial) | 11 GB Laptop or phone | |
| OmniVoice Xiaomi | 613M | – | CC BY-NC (Non-commercial) | 3.75 GB Laptop or phone |