Superwhisper introduces s1-mini: open-weights text normalizer for clean asr output

Superwhisper, a company specializing in speech-to-text and text normalization technologies, has launched S1-mini, an open-weights text normalizer designed to transform raw automatic speech recognition (ASR) transcripts into polished written text. The model, available on Hugging Face under an Apache 2.0 license with a naming clause, operates as a text normalizer rather than a transcriber or chat model. It is intended to be used after ASR systems, such as Whisper or Parakeet, to refine raw transcripts by removing filler words, resolving self-corrections, applying punctuation and capitalization, and converting spoken numbers, dates, and email addresses into written form. S1-mini is fine-tuned from Qwen/Qwen3-0.6B and has 596M unique parameters, with 28 layers and support for GQA. The model is limited to English in its first release and requires a fixed system prompt, a control line, and a raw transcript for input. It supports four styling options, two structural formats, and two contextual settings, with all combinations trained. Superwhisper reports a token accuracy of 94.8% on a held-out set of 7,519 cases, with a text-edit error rate of 11.6%. The model also performs well in email formatting, identifying greeting lines 99.3% of the time and sign-offs 97.9% of the time. It matches the correct output structure 97.6% of the time and generates exact email addresses in 92% of cases. The model is constrained in its functionality, as it does not add content, correct facts, or rewrite dialect. It also returns an empty string for filler-only input. S1-mini requires specific configuration settings, including disabling thinking mode and decoding greedily, to ensure optimal performance. The model is part of the S1 family, which includes S1-Voice, a cloud-based speech-to-text model, and S1-Language, an instruction-following model for cleanup and formatting. S1-Voice achieves transcription speeds up to 46 times faster than speaking time, with a word error rate of 6.8% across eight datasets, including LibriSpeech, where it drops to 2.2%. S1-Language is available alongside models from Anthropic, OpenAI, and Groq in the model picker. Superwhisper also highlights the importance of using the correct input format and configuration to avoid output degradation or garbling. The company encourages users to explore the model weights and technical details for further insights.