|

Meet S1-mini: Superwhisper’s 462 MB Open-Weights Text Normalizer That Turns Raw ASR Transcripts Into Clean Written Text

ByRicardo

Superwhisper has launched the S1 family of models: S1-Voice, S1-Language, and S1-mini. S1-Voice is a cloud speech-to-text mannequin, and S1-Language is a cloud instruction-following mannequin for cleanup and formatting. The one that’s fairly fascinating exterior the app is S1-mini, launched with open weights on Hugging Face. S1-mini is a 0.6B textual content normalizer, not a transcriber and never a chat mannequin. It sits after computerized speech recognition and rewrites uncooked transcripts as clear written textual content: fillers eliminated, self-corrections resolved to what the speaker landed on, punctuation and capitalization utilized, and spoken numbers, dates, forex, and e-mail addresses rendered in written type. It is fine-tuned from Qwen/Qwen3-0.6B, covers English solely in launch v1, and is steered completely by a three-axis management line positioned above the transcript. Superwhisper studies 94.8% token accuracy on a held-out set of seven,519 instances, measured grasping on the quantized construct.

Is it deployable?

Yes, however solely S1-mini. S1-mini is printed on Hugging Face below Apache 2.0 plus a naming clause. S1-Voice and S1-Language are Superwhisper-hosted companies, so they’re consumable, not self-hostable.

  • Company degree: Any degree. The Q4_K_M GGUF build is a 462 MB file that runs on a laptop computer CPU. Solo builders can ship it inside a desktop app. Enterprises can run it behind a VPC the place audio transcripts can’t go away the community.
  • Industries: Healthcare and medical documentation, authorized, monetary companies, buyer assist, developer tooling, accessibility and reside captioning.
  • Applications: Dictation apps, meeting-notes instruments, reside captioning, voice-driven editors, voice-to-CRM entry, and any pipeline that turns uncooked ASR output into textual content a human will learn.

What S1-mini does

S1-mini is a textual content normalizer, not a transcriber and never a chat mannequin. It sits after computerized speech recognition:

audio → ASR (Whisper, Parakeet, …) → S1-mini → clear textual content

It removes filler phrases, resolves false begins and self-corrections to the worth the speaker landed on, applies punctuation and capitalization, and renders spoken numbers, dates, instances, forex, and e-mail addresses in written type. Say “assist at superwhisper dot com” and also you get [email protected].

The mannequin is fine-tuned from Qwen/Qwen3-0.6B. It has 596M distinctive parameters (0.44B non-embedding), 28 layers, 16 question heads and eight key/worth heads with GQA, and BF16 weights. The Hub sidebar studies 0.8B as a result of the tied embedding is saved twice; the cardboard explains the discrepancy explicitly. Release v1 covers English solely, and beneficial enter is roughly 1,000 tokens.

The management line is your entire interface

S1-mini takes a hard and fast system immediate, then a management line, then the uncooked transcript:

Copy Code
[Styling: <value>] [Structure: <value>] [Context: <value>]
<uncooked transcript>

Styling takes informal, semi-casual, semi-formal, or formal. Structure takes prose or lists. Context takes basic or e-mail. All three axes are impartial, and each mixture was skilled. Send values exterior these units, or reword the system immediate, and output can degrade or garble. Note the small mismatch value realizing: the Superwhisper app exposes a five-stop tone slider that provides a “balanced” preset, whereas the open weights doc 4 skilled Styling values.

The mannequin can be constrained by design. It doesn’t add content material you didn’t say, right details, soften profanity, or rewrite dialect. Filler-only enter returns an empty string, and integrations ought to deal with that as a sound consequence.

Two settings that break most integrations

First, enable_thinking=False is required. The chat template is Qwen3’s, unchanged, and Qwen3 defaults to considering on. S1-mini was skilled with considering off, so the assistant flip should open with an empty <assume> block. Omit the flag and also you normally get no usable output in any respect.

Second, decode greedily. generation_config.json ships do_sample: false. The GGUF builds nonetheless carry Qwen3’s inherited temp = 0.6, top_p = 0.95, and top_k = 20 metadata, so move temperature 0 explicitly on each request. In llama.cpp, use --jinja with --chat-template-kwargs '{"enable_thinking":false}' moderately than --reasoning-budget 0, which degrades output.

Reported analysis

Superwhisper evaluated S1-mini on a held-out set of seven,519 instances throughout 104 transcripts. Token accuracy is 94.8%, measured grasping on the Q4_K_M construct, with a text-edit error price of 11.6%. On email-formatted textual content it identifies the greeting line 99.3% of the time and the sign-off 97.9%. It matches the right output construction, listing versus paragraph, 97.6% of the time, and produces precise e-mail addresses in 92% of instances. Fewer than 1% of generations present looping or truncation, and the mannequin appropriately withholds output 98.6% of the time when nothing needs to be transcribed. These are vendor-reported numbers on an inner take a look at set, not third-party outcomes.

The two cloud fashions

S1-Voice is the hosted speech-to-text mannequin. Superwhisper studies transcription as much as 46x sooner than talking time, with most dictations below 30 seconds showing 0.32 seconds after you cease. Across eight datasets together with assembly audio and earnings calls, it averages 6.8% phrase error price and drops to 2.2% on LibriSpeech. Superwhisper says that 6.8% common was the bottom of 15 fashions it examined, and that S1-Voice scored 83 out of 100 on its blended metric towards WisprFlow’s 76.

S1-Language is the hosted instruction-following mannequin for cleanup, formatting, and summarization, and it seems within the mannequin picker alongside fashions from Anthropic, OpenAI, and Groq. The beneficial defaults are Cohere Transcribe plus S1-mini offline, or S1-Voice plus S1-Language within the cloud.

Interactive explainer

The embed beneath allows you to change every control-line axis and watch the output change. Every enter/output pair is taken verbatim from the mannequin card.


Key Takeaways

  • S1-mini is a 0.6B open-weights textual content normalizer for ASR output, not a transcriber or chat mannequin.
  • The Q4_K_M GGUF is 462 MB and runs on a laptop computer CPU, so on-device deployment is life like.
  • A hard and fast system immediate plus a three-axis management line is the one steering mechanism.
  • enable_thinking=False and temperature 0 are necessary; most integration bugs hint to those.

Check out the Model Weights and Technical Details. Also, be at liberty to comply with us on Twitter and don’t overlook to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to accomplice with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and so forth.? Connect with us

The submit Meet S1-mini: Superwhisper’s 462 MB Open-Weights Text Normalizer That Turns Raw ASR Transcripts Into Clean Written Text appeared first on MarkTechPost.

Similar Posts