AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
Large Language Models

Superwhisper Releases S1-mini: A 462 MB Open-Weights Text Normalizer for ASR Transcripts

Superwhisper introduced S1-mini, a compact open-weights model that cleans raw ASR transcripts into readable text, runs on a laptop CPU, and achieves 94.8% token accuracy.

AI News Feed · 2026-08-21 23:37

S1-mini is not a transcriber or a chat model; it sits after an ASR system such as Whisper or Parakeet. It removes filler words, resolves self-corrections to the speaker's final intent, applies punctuation and capitalization, and converts spoken numbers, dates, currency, and email addresses into written form. For example, saying "support at superwhisper dot com" produces support@superwhisper.com. The model is fine-tuned from Qwen/Qwen3-0.6B and covers English only in its first release.

The model contains 596 million unique parameters (0.44 billion non-embedding), uses 28 layers with grouped query attention, and is steered by a three-axis control line that adjusts styling (casual to formal), structure (prose or lists), and context (general or email). Superwhisper notes that all combinations of these attributes were trained, and sending values outside the defined sets can degrade output. The release also highlights a mismatch: the Superwhisper app exposes a five-stop tone slider, while the open weights document four trained styling values.

Superwhisper reports evaluation results on a held-out set of 7,519 cases across 104 transcripts. The model achieves 94.8% token accuracy measured greedily on the Q4_K_M quantized build, with a text-edit error rate of 11.6%. It correctly identifies email greeting lines 99.3% of the time and sign-offs 97.9%, matches the correct output structure 97.6% of the time, and produces exact email addresses in 92% of cases. Fewer than 1% of generations show looping or truncation, and the model withholds output correctly 98.6% of the time when nothing should be transcribed. These figures are vendor-reported on an internal test set, not third-party results.

Deployment requires specific settings: the model must be used with enable_thinking=False because Qwen3's default thinking mode is enabled, and decoding should be greedy with temperature 0. The model is constrained by design—it does not add content, correct facts, soften profanity, or rewrite dialect. Filler-only input returns an empty string, which should be treated as a valid result.

S1-mini is intended for use in dictation apps, meeting notes, live captioning, voice-driven editors, and any pipeline that converts raw ASR output into human-readable text. The two related cloud models, S1-Voice and S1-Language, remain hosted services and are not self-hostable.