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

- Superwhisper dropped S1-mini — a 462 MB open-weights text normalizer that sits after ASR and turns your rambling voice garbage into something a human can actually read. It strips fillers, resolves self-corrections, formats emails and dates, all on a laptop CPU. 94.8% token accuracy on their own test set, which is vendor-reported, so file that under "trust but verify the PoD seal." The catch? It's English-only, it won't fix your facts, and if you feed it pure filler it returns nothing — which is honestly the most honest thing an AI model has done this year. Apache 2.0 with a naming clause means you can self-host it behind a VPC where your audio transcripts can't escape into some data terrorist's pipeline. Meanwhile S1-Voice and S1-Language stay locked behind Superwhisper's cloud — because of course the good stuff stays hosted. For anyone building dictation tools or meeting-note pipelines, this is a genuinely useful 0.6B model that does one job and doesn't pretend to be a chatbot. Where's my cut?