|
One more thing…
Superwhisper · Post-ASR · Aug 20
Meet S1-mini: Superwhisper’s 462 MB Open-Weights Text Normalizer That Turns Raw ASR Transcripts Into Clean Written Text
You already know this problem. The transcript is accurate and unusable. Every “um” is there, the false start is there, and the sentence where the speaker said Alice and then corrected themselves to Bob is faithfully preserved in both versions. So you pipe it through a frontier model to clean it up, and now every dictation costs an API call and leaves the machine.
S1-mini is that stage as a dedicated 0.6B model. It does not transcribe and it does not chat. It removes fillers, resolves self-corrections to what the speaker landed on, applies punctuation and capitalization, and renders spoken numbers, dates, currency and email addresses in written form. It is constrained by design: it will not add content, correct facts, soften profanity or rewrite dialect, and filler-only input returns an empty string. Steering is one control line above the transcript with three independent axes, styling, structure and context, and every combination was trained.
The quantized build is a 462 MB file that runs on a laptop CPU, published under Apache 2.0 with a naming clause and fine-tuned from Qwen3-0.6B. Superwhisper reports 94.8% token accuracy on a held-out set of 7,519 cases, an 11.6% text-edit error rate, and correct withholding 98.6% of the time when nothing should be transcribed — vendor numbers on an internal set, English only in v1. Two settings decide whether it works at all: thinking must be disabled, and you must pass temperature 0 explicitly, because the quantized builds still carry Qwen3’s inherited sampling metadata.
Read on Marktechpost →
◆ Model card
◆ GGUF build
◆ Technical write-up
|
