Meet S1-mini: Superwhisper's 462 MB Open-Weights Text Normalizer
Discover S1-mini, a 462 MB open-weights text normalizer fine-tuned from Qwen3-0.6B by Superwhisper, designed to turn raw ASR transcripts into clean written text with 94.8% token accuracy.
Stock photo for illustration only, not from the actual event
- S1-mini is a 462 MB open-weights text normalizer built to clean up raw speech-to-text transcripts.
- Fine-tuned from Qwen/Qwen3-0.6B with 596M total parameters, supporting English only in release v1.
- Achieves a token accuracy of 94.8% based on evaluations across an internal test set of 7,519 cases.
- Published on Hugging Face under the Apache 2.0 license with an additional naming clause.
In the field of artificial intelligence and speech processing, Superwhisper has introduced S1-mini, a 462 MB open-weights model functioning as a text normalizer. Rather than acting as a transcriber or a chat model, S1-mini is positioned directly after automatic speech recognition systems to turn raw, unformatted transcripts into clean, readable written text.
The processing pipeline places the model right after ASR tools such as Whisper or Parakeet, following a clear sequence from audio input to ASR output, and subsequently through S1-mini to produce polished written text ready for practical use.
Deploying a dedicated text normalizer downstream from ASR engines effectively addresses common issues in raw speech transcripts, such as filler words, irregular spacing, and informal structures. Utilizing a lightweight specialized model like S1-mini offers an efficient alternative to heavy general-purpose language models for formatting tasks.
Technically, S1-mini is fine-tuned from Qwen/Qwen3-0.6B. It features 596 million unique parameters (0.44B non-embedding), 28 layers, 16 query heads, and 8 key/value heads with GQA, along with BF16 weights. The Hugging Face sidebar reports 0.8B because the tied embedding is stored twice, a discrepancy explicitly explained in the model card. Release v1 covers English only, with a recommended input size of roughly 1,000 tokens.
Stock photo for illustration only, not from the actual event
The model is intentionally constrained by design to maintain reliability: it does not add unstated content, correct facts, soften profanity, or rewrite dialects. Input consisting solely of fillers returns an empty string, which integrated applications should handle as a valid outcome.
Superwhisper evaluated S1-mini on a held-out set of 7,519 cases across 104 transcripts, yielding specific benchmark metrics:
- Token accuracy reaches 94.8%, measured greedy on the Q4_K_M build.
- Text-edit error rate stands at 11.6%.
- Email greeting lines are identified 99.3% of the time, and sign-offs 97.9%.
- Correctly matches output structures between lists and paragraphs in 97.6% of cases.
- Produces exact email addresses in 92% of instances.
- Fewer than 1% of generations exhibit looping or truncation.
- Correctly withholds output 98.6% of the time when no transcription is required.
These figures represent vendor-reported metrics on an internal test set rather than third-party validations. Additionally, the broader Superwhisper ecosystem includes S1-Voice, a hosted speech-to-text model operating up to 46x faster than speaking time, and S1-Language, a hosted instruction-following model dedicated to cleanup, formatting, and summarization.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment