Superwhisper launches S1-mini to clean up ASR transcripts

Superwhisper has introduced the S1 family of models, highlighted by the open-weights release of S1-mini on Hugging Face under an Apache 2.0 license with a naming clause. S1-mini is a 0.6-billion-parameter text normalizer designed to sit directly after automatic speech recognition (ASR) systems like Whisper or Parakeet. Rather than transcribing audio or acting as a conversational agent, the model processes raw transcripts to remove filler words, resolve self-corrections, apply punctuation, and format spoken elements like emails and currencies. The Q4_K_M GGUF build is a 462 MB file that runs locally on a laptop CPU, making it highly accessible for desktop applications and secure enterprise environments.

Fine-tuned from Qwen/Qwen3-0.6B, S1-mini features 596 million unique parameters, 28 layers, 16 query heads, and 8 key/value heads with Grouped-Query Attention. The model supports English-only inputs of up to roughly 1,000 tokens. Developers steer the model using a three-axis control line specifying styling (casual, semi-casual, semi-formal, or formal), structure (prose or lists), and context (general or email). Superwhisper notes that integrations must explicitly set "enable_thinking" to false and force greedy decoding with a temperature of zero to prevent output degradation.

In internal evaluations across a held-out set of 7,519 cases from 104 transcripts, S1-mini achieved a 94.8% token accuracy and an 11.6% text-edit error rate. It successfully identified email greeting lines 99.3% of the time, matched correct output structures 97.6% of the time, and generated exact email addresses with 92% accuracy. Fewer than 1% of generations showed looping or truncation, and the model correctly withheld output 98.6% of the time when nothing was spoken. Alongside S1-mini, Superwhisper launched two hosted cloud services: S1-Voice, a speech-to-text model boasting a 6.8% average word error rate and a 2.2% rate on LibriSpeech, and S1-Language, an instruction-following model. S1-Voice transcribes up to 46 times faster than speaking speed, delivering short dictations in 0.32 seconds and scoring 83 out of 100 on Superwhisper's blended metric compared to WisprFlow's 76.