Fish Audio S1-Mini: What It Is and How to Actually Try It

S1-mini is Fish Audio's 0.5 billion parameter text-to-speech model, a distilled sibling of the 4B S1 flagship that topped the HuggingFace TTS-Arena-V2 leaderboard. There is no documented local install path right now, so here is what the model actually is and where you can hear it.

6 minBeginner

Fish Audio S1-mini is a 0.5 billion parameter open-weight text-to-speech model, released May 31, 2025 as the distilled, lightweight sibling of Fish Audio's 4B-parameter S1 flagship. It was trained on over 2 million hours of multilingual audio, covers 13 languages, and supports dozens of inline emotion and tone tags. The weights are on Hugging Face under a non-commercial license, but Fish Audio's own GitHub repo has since moved on to document its newer S2 Pro model instead, so there is no official local inference script for S1-mini right now.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts and licensing against the primary source before recommending an install path, and I say so plainly when that path does not exist yet.

S1-mini entered the top 20 of Hugging Face's text-to-speech trending list. At time of writing it has 732 likes and 1,944 downloads on the fishaudio/s1-mini model page, and the linked demo Space, fishaudio/s1-mini, has 697 likes of its own. That Space is the fastest way to hear the model without downloading anything.

Specs at a glance

Field	Value
Publisher	Fish Audio
Parameters	0.5 billion (distilled from the 4B S1 flagship)
Training data	2M+ hours of multilingual audio
Pipeline	Text to speech with inline emotion, tone and sound-effect tags
Languages	13: English, Chinese, Japanese, German, French, Spanish, Korean, Arabic, Russian, Dutch, Italian, Polish, Portuguese
License	CC-BY-NC-SA-4.0 (non-commercial)
Release date	May 31, 2025
Local inference	Not documented; the linked GitHub repo now covers S2 Pro instead

Non-commercial license, and it's gated

S1-mini is gated on Hugging Face. You have to accept Fish Audio's usage terms before downloading model.pth, codec.pth, and the tokenizer files. The CC-BY-NC-SA-4.0 license covers research and personal projects; it does not cover commercial use.

How it scores against the flagship

Fish Audio's announcement post says the S1 family ranked #1 in Human Subjective Evaluation on Hugging Face's TTS-Arena-V2, a community leaderboard where listeners vote on which model's audio sounds more natural. S1-mini is the compressed version built for that same quality bar at a fraction of the size: on Fish Audio's own English benchmark it posts a 0.011 word error rate and 0.005 character error rate, against 0.008 and 0.004 for the full S1. That is a small quality gap for an 8x reduction in parameter count.

Metric (English)	S1 (4B)	S1-mini (0.5B)
Word error rate	0.008	0.011
Character error rate	0.004	0.005

Inline emotion and tone control

Like the rest of the Fish Speech line, S1-mini does not rely on a fixed list of speaker presets. You write bracket tags directly into the text you want spoken, and the model renders them as part of the delivery. The model card lists dozens of emotion tags such as (angry), (excited), (nervous), and (sarcastic), plus a shorter set of tone tags, (shouting), (whispering), (soft tone), (screaming), and (in a hurry tone), and a handful of non-speech sound tags like (laughing), (sighing), and (crying loudly).

Why there's no install section here

Fish Audio's fish-speech GitHub repo (32,900+ stars) is the link on S1-mini's model card, but its README has been rewritten to document S2 Pro, the newer 4.56B-parameter flagship the company shipped afterward. It does not mention S1 or S1-mini at all anymore. The S1-mini model card itself has no usage, quick start, or inference section, and its files are plain PyTorch checkpoints (model.pth, codec.pth) rather than a format a standard transformers pipeline loads out of the box. We are not going to invent install steps that no official source documents.

Your two real options today

Try the model for free in the browser via the fishaudio/s1-mini Space on Hugging Face, or use Fish Audio's hosted API through the Fish Audio Playground if you need it in a pipeline. If Fish Audio publishes dedicated S1-mini inference code later, we'll add a local-install section here.

FAQ

What is Fish Audio S1-mini?

A 0.5 billion parameter open-weight text-to-speech model from Fish Audio, distilled from the company's larger 4B S1 model. It covers 13 languages and supports inline emotion and tone tags written directly into the input text.

Can I run S1-mini locally right now?

Not with an official, documented path. The weights are downloadable from Hugging Face once you accept the gated license, but Fish Audio's GitHub repo now documents its newer S2 Pro model instead, and the S1-mini model card has no quick start or inference instructions.

Is S1-mini free to use commercially?

No. It ships under CC-BY-NC-SA-4.0, which covers research and non-commercial projects only.

How does it compare to the full S1 model?

S1-mini trades size for a small accuracy drop: 0.011 word error rate versus 0.008 for S1, and 0.005 character error rate versus 0.004, on Fish Audio's English benchmark. At 0.5B parameters it is roughly an eighth the size of the 4B S1.

Where can I hear a sample?

The fishaudio/s1-mini Space on Hugging Face runs the model in your browser with no setup. Fish Audio also offers a hosted Playground on fish.audio if you want to generate audio through an API.

Where to go from here

Need a TTS model you can actually self-host today? Run Kokoro-82M locally covers a smaller, Apache-licensed model with a working install path.
Want Fish Audio's documented flagship instead? Run Fish Audio S2 Pro locally covers the model that replaced S1 in the company's own GitHub repo.
Looking for a model with commercial-friendly licensing? Run Qwen3-TTS-12Hz-1.7B-VoiceDesign locally covers an alternative voice model you can install today.
Compare S1-mini against other open TTS models: browse the local AI model directory to see specs side by side.

Local models run better with more VRAM. CompareRTX GPUs on Amazonbefore you upgrade.(affiliate link. We may earn a commission at no extra cost. Disclosure)