**Superwhisper S1‑mini: A Compact, Open‑Weight Text Normalizer for ASR**
Superwhisper has introduced the S1 family of models, among which S1‑mini stands out as a fully open‑weight solution designed specifically to clean up automatic speech recognition (ASR) output. Unlike S1‑Voice and S1‑Language, which are hosted services, S1‑mini is published on Hugging Face under Apache 2.0 (plus a naming clause) and can be self‑hosted, for example in offline desktop apps or enterprise environments where transcripts must remain on‑premise.
**What S1‑mini Does**
S1‑mini is a text normalizer, not a transcriber and not a chat model. It takes raw ASR transcripts as input and outputs clean, written text. Its responsibilities include removing fillers, resolving self‑corrections to the speaker’s final wording, applying correct punctuation and capitalization, and rendering spoken numbers, dates, currencies, and email addresses in proper written form. For example, the phrase “support at superwhisper dot com” becomes [email protected].
Technically, S1‑mini is fine‑tuned from Qwen/Qwen3‑0.6B. It contains 596M unique parameters (0.44B non‑embedding), spans 28 layers, and uses GQA with 16 query heads and 8 key/value heads, all in BF16. The Hugging Face metadata reports 0.8B total due to the tied embedding storage. The model is English‑only in its current release and recommends inputs of roughly 1,000 tokens.
**Control‑Line Interface**
Interaction with S1‑mini is entirely driven by a fixed system prompt and a single control line placed above the transcript. This control line defines three independent axes:
– **Styling**: casual, semi‑casual, semi‑formal, or formal
– **Structure**: prose or lists
– **Context**: general or email
Every combination of these values has been trained. Users must keep the control line in one of the supported sets; otherwise output quality may degrade or produce garbled text. Importantly, the model does not add content that was not spoken, correct facts, soften profanity, or rewrite dialects. If the input contains only fillers, the model returns an empty string, and downstream integrations should treat this as a valid result.
**Integration Requirements and Pitfalls**
Two settings are critical for successful integration:
1. **`enable_thinking=False`** must be set. S1‑mini was trained with Qwen3’s default “thinking” mode turned off, so the model must start with an empty `
2. **Decoding must be greedy**. The reference `generation_config.json` specifies `do_sample: false`. While GGUF builds still carry Qwen3 metadata with temperature 0.6, top‑p 0.95, and top‑k 20, users should explicitly pass temperature 0 to avoid sampling artifacts. In llama.cpp, the recommended invocation uses `–jinja` with `–chat-template-kwargs ‘{“enable_thinking”:false}’` rather than `–reasoning-budget 0`.
**Performance Evaluation**
Superwhisper evaluated S1‑mini on a held‑out set of 7,519 cases across 104 transcripts. Token accuracy reached 94.8% when decoded greedily on the Q4_K_M quantized build, with a text‑edit error rate of 11.6%. On email‑formatted text, the greeting line was identified 99.3% of the time and the sign‑off 97.9%. The model matched the correct output structure (list vs. paragraph) 97.6% of the time and produced exact email addresses in 92% of cases. Fewer than 1% of generations showed looping or truncation, and the model correctly withheld output 98.6% of the time when the input contained nothing worth transcribing. These results are based on internal testing and should be considered vendor‑reported rather than third‑party verified.
**Deployment Scenarios**
Because the Q4_K_M GGUF build is under 462 MB and runs on laptop CPUs, S1‑mini is suitable for on‑device use. Organizations can deploy it behind a VPC to ensure that sensitive audio transcripts never leave their network. Recommended applications include dictation apps, meeting‑notes tools, live captioning, voice‑driven editors, voice‑to‑CRM pipelines, and any workflow that converts raw ASR output into human‑readable text.
**Key Takeaways**
– S1‑mini is a 0.6B open‑weight text normalizer for ASR output, not a transcriber or chat model.
– The Q4_K_M GGUF is 462 MB and runs on a laptop CPU, making on‑device deployment feasible.
– A fixed system prompt plus a three‑axis control line is the sole mechanism for steering the model.
– `enable_thinking=False` and greedy decoding with temperature 0 are mandatory; most integration issues stem from neglecting these settings.
For more information, users can consult the [Superwhisper blog](https://superwhisper.com/blog/s1), the [s1‑mini model card](https://huggingface.co/superwhisper/s1-mini), and the [GGUF builds page](https://huggingface.co/superwhisper/s1-mini-GGUF). The model weights and technical details are also available for those who wish to integrate S1‑mini into their own tools and pipelines.
