Superwhisper Launches 462 MB Open-Weights Text Normalizer

Automatic speech recognition systems typically produce literal transcripts filled with verbal fillers, repetitions, false starts and inconsistent punctuation, leaving substantial editing work before the text is suitable for publication or business use. Voice-application developer Superwhisper is targeting that post-processing bottleneck with a dedicated text-normalization model designed for privacy-sensitive and low-latency uses such as meeting notes, voice input and offline transcription.

Superwhisper has released S1-mini, an open-weights model fine-tuned from Qwen3-0.6B that occupies just 462 MB and is designed to run smoothly on a laptop CPU. The model converts raw ASR output into polished written text by removing filler words, correcting speech errors and inserting precise punctuation. Local execution could reduce reliance on cloud processing while giving developers a compact component for cleaning transcripts inside voice-enabled applications.