- Published on
Superwhisper’s 462 MB Open-Weights Text Normalizer That Turns Raw ASR Transcripts Into Clean Written Text
- Authors
- Name
- DeepHorus
Superwhisper, an AI company focused on speech-to-text technology, has announced the release of its S1 model family, including the open-weights text normalizer S1-mini. This lightweight model, available on Hugging Face, is designed to clean up raw transcripts from automatic speech recognition (ASR) systems like Whisper. It transforms messy spoken text into polished written text, marking a significant step for developers building voice-to-text pipelines.
The S1-mini model is a specialized 0.6 billion parameter text normalizer, not a transcriber or chat model. It is fine-tuned from the Qwen/Qwen3-0.6B base model and focuses exclusively on English for its initial release. Its core function is post-processing: it removes filler words, resolves self-corrections, applies proper punctuation and capitalization, and formats elements like numbers, dates, currency, and email addresses correctly.
Technically, S1-mini is remarkably efficient. The model's quantized Q4_K_M GGUF build is just 462 MB, making it deployable on a single laptop CPU. This small footprint allows for widespread use cases, from individual developers integrating it into desktop applications to enterprises running it within secure, private networks where data cannot leave the premises.
The model's interface is controlled by a simple three-axis prompt specifying styling, structure, and context. Users can choose from casual to formal tones, prose or list outputs, and general or email contexts. Superwhisper reports a high 94.8% token accuracy on a test set of 7,519 cases when using greedy decoding, which is the required setting for optimal performance.
This release is particularly impactful for industries where accurate documentation from audio is critical. Potential applications include healthcare and clinical documentation, legal transcription, financial services, customer support logs, and accessibility tools like live captioning. It addresses the common problem of ASR output being too raw for direct human reading.
The S1-mini model is released under the permissive Apache 2.0 license, though it includes a naming clause. It is part of a broader family that includes cloud-based S1-Voice for transcription and S1-Language for instruction following, which are hosted services. Future updates are expected to expand language support beyond the current English-only version 1.
