S1-mini: A 462 MB Open-Weights Model That Cleans Raw ASR Transcripts

Superwhisper has published S1-mini on Hugging Face under Apache 2.0: a 462 MB, 0.6B-parameter text normalizer that sits after an ASR model and rewrites raw transcripts into clean written text. It strips fillers, resolves self-corrections, applies punctuation and capitalization, and converts spoken numbers, dates, currency, and email addresses into their written forms. The quantized GGUF build runs on a laptop CPU, and Superwhisper reports 94.8% token accuracy on an internal test set of 7,519 cases.

What happened

Detail	Value
Model	S1-mini (open weights on Hugging Face)
Base model	Qwen/Qwen3-0.6B
Parameters	596M unique (0.44B non-embedding), 28 layers
Quantized file size	462 MB (Q4_K_M GGUF)
Token accuracy (internal)	94.8% on 7,519 held-out cases
Text-edit error rate	11.6%
License	Apache 2.0 plus a naming clause
Language coverage	English only (v1)
Recommended input length	~1,000 tokens

Superwhisper released three models at once: S1-Voice (a hosted speech-to-text model), S1-Language (a hosted cleanup and formatting model), and S1-mini. Only S1-mini ships with open weights. The other two are Superwhisper-hosted services you consume through the app, not self-host.

S1-mini is not a transcriber and not a chat model. It plugs into the end of an ASR pipeline: audio goes into Whisper, Parakeet, or any other recognizer, and then S1-mini takes that raw output and rewrites it. Say “support at superwhisper dot com” to your microphone and the model emits [email protected]. Filler-only input returns an empty string, which the model card says integrations should treat as a valid result, not an error.

How the control line works

The entire interface is a single line placed above the raw transcript in the prompt:

[Styling: <value>] [Structure: <value>] [Context: <value>]

Each axis accepts a fixed set of values:

• Styling: casual, semi-casual, semi-formal, or formal
• Structure: prose or lists
• Context: general or email

All combinations were trained. Sending anything outside those sets, or rewording the system prompt, can produce garbled output. One small inconsistency worth knowing: the Superwhisper app shows a five-stop tone slider that includes a “balanced” preset, while the open-weights card documents only four Styling values.

Two integration mistakes that will break your build

The model card flags both of these as common failure points, and skipping either one typically produces no usable output at all.

1. Set enable_thinking=False. S1-mini uses Qwen3’s unchanged chat template. Qwen3 defaults to thinking mode on, but S1-mini was trained with it off. The assistant turn must open with an empty <think> block, and the flag enforces that.
2. Decode greedily (temperature 0). The generation_config.json ships do_sample: false, but the GGUF builds inherit Qwen3 metadata with temperature 0.6, top_p 0.95, and top_k 20. Pass temperature 0 explicitly on every request. In llama.cpp, use –jinja with –chat-template-kwargs rather than –reasoning-budget 0, which the card says degrades output.

What the evaluation numbers actually say

Superwhisper ran S1-mini against a held-out set of 7,519 cases across 104 transcripts. These are vendor-reported figures on an internal test set, not third-party benchmarks. With that caveat stated, the results are specific enough to be useful:

• Token accuracy: 94.8%, measured greedy on the Q4_K_M build
• Text-edit error rate: 11.6%
• Greeting-line identification on email-formatted output: 99.3%
• Sign-off identification: 97.9%
• Correct output structure (list vs. paragraph): 97.6%
• Exact email address reproduction: 92%
• Looping or truncation in fewer than 1% of generations
• Correctly withholds output when nothing should be transcribed: 98.6%

The two hosted models for context

S1-Voice, the cloud ASR model, reports 46x faster than speaking time for short dictations, averaging 0.32 seconds latency on clips under 30 seconds. Superwhisper says S1-Voice averaged 6.8% word error rate across eight datasets including meeting audio and earnings calls, and 2.2% on LibriSpeech specifically. The company says that 6.8% figure was the lowest of 15 models it tested internally, and that S1-Voice scored 83 out of 100 on its blended metric versus WisprFlow’s 76. S1-Language, the hosted cleanup model, appears in the app alongside models from Anthropic, OpenAI, and Groq.

Why it matters

Most ASR pipelines stop at the transcript. Raw output from Whisper or similar models is accurate but messy: fillers, incomplete sentences, spoken numbers. Getting that text into a CRM, a clinical note, or a customer support ticket requires a cleanup step. S1-mini formalizes that step as a small, offline, licensable model anyone can run without sending audio data to a third-party server.

The 462 MB size matters for deployment. A solo developer can bundle it inside a desktop app. A healthcare or legal team can run it behind a VPC where audio transcripts never leave the network. That is a meaningful distinction from cloud-only alternatives, especially in regulated industries where data residency requirements are strict. If you are already building voice-driven tools or voice-to-CRM workflows, this fits directly into the normalization layer.

Our take

The framing is honest: this is a normalizer, not a general assistant. That specificity is a feature. Models that try to do everything tend to do nothing well at a small parameter count. Superwhisper kept the scope tight, trained every control-line combination, and published the failure modes clearly in the card. The 94.8% token accuracy figure is vendor-measured on an internal set, so treat it as directional rather than definitive. The 11.6% text-edit error rate is the number we would watch more closely in production.

The two integration gotchas (thinking mode and temperature) are the kind of thing that wastes a full afternoon if you do not read the card first. Read the card first. If you are evaluating this for a client pipeline, the GGUF build is small enough to test locally in an hour. We would not put it in front of a clinical documentation workflow without running your own evaluation set against the specific domain vocabulary you care about.

For teams already using workflow automation to route voice input into downstream systems, S1-mini slots in as a clean post-processing node without adding a cloud dependency. That is worth a quick proof of concept before committing to a hosted solution.

Source: Marktechpost