S1-mini is a 0.6B parameter text normalizer released by Superwhisper with open weights on Hugging Face. It is not a speech recognizer or chat model; it operates on raw ASR transcripts to produce polished written text by removing fillers, fixing self‑corrections, adding punctuation/capitalization, and converting spoken numbers, dates, currency, and email addresses to their written forms.
The model is a fine‑tuned version of Qwen/Qwen3-0.6B, featuring 596M unique parameters (0.44B non‑embedding), 28 transformer layers, 16 query heads and 8 key/value heads using grouped‑query attention, and BF16 weights. It accepts a fixed system prompt followed by a three‑axis control line — [Styling: <value>] [Structure: <value>] [Context: <value>] — then the raw transcript. Styling chooses casual, semi‑casual, semi‑formal, or formal; Structure selects prose or lists; Context picks general or email. All combinations were trained.
- Parameter count: 596M unique (0.44B non‑embedding)
- Layers: 28; Heads: 16 query / 8 KV (GQA)
- Precision: BF16
- Context window: ~1,000 tokens recommended
- License: Apache 2.0 + naming clause
- Language: English (v1)
- Reported token accuracy: 94.8% on 7,519‑case held‑out set (greedy, quantized)
Why this matters
The availability of S1-mini under Apache 2.0 lets developers embed a deterministic, size‑constrained normalizer directly into on‑device or private‑cloud ASR pipelines, avoiding latency and data‑privacy costs of remote cleanup services. Its three‑axis control interface offers reproducible styling without retraining, while the reported 94.8% token accuracy indicates a strong baseline for post‑processing English transcripts.
Share this article
Found this insightful? Share it with your community on Reddit, X, or copy the link.