Speech recognition systems frequently produce lowercase, unpunctuated, and speaker‑style transcripts that are hard for humans to read. S1-mini targets that exact gap: a compact, deterministic text-normalizer trained to map noisy ASR output into readable, display-ready text while preserving the speaker’s intended final phrasing.
What Sets It Apart
- Task-focused fine-tune: trained specifically to perform a single deterministic transformation (raw ASR → cleaned text). This reduces hallucination risk and makes outputs predictable for UI pipelines.
- Small and deployable: based on Qwen/Qwen3-0.6B with ~596M unique parameters and BF16 weights; quantized GGUF builds (~462 MiB) run comfortably on laptop CPUs, enabling on-device dictation and live captioning.
- Precise control surface: inputs must include a fixed system prompt plus a control line with three axes (Styling, Structure, Context). This explicit control yields consistent register (casual → formal), list handling, and email layout without prompt engineering variance.
- Evaluation-backed: measured 94.8% token accuracy on a held-out English set of 7,519 cases covering numbers, self-corrections, lists, email style and adversarial inputs, so it performs reliably on typical dictation errors.
How it works (brief)
S1-mini is a causal language model fine-tuned from Qwen3-0.6B to perform normalization rather than open-ended generation. The runtime contract is strict: the model expects the system prompt, then a control line, a newline, and one raw transcript. It was trained with the chat template's empty <think> block disabled, so integrations must preserve the exact template and set enable_thinking=false (or equivalent) to get usable output.
Who should use it and trade-offs
Great fit if you need: on-device or low-cost server post-processing for ASR; deterministic, display-ready transcripts for dictation apps, meeting notes, live captions, or email drafting; a model that enforces consistent styling via a simple control line. Look elsewhere if you need: multilingual normalization (S1-mini is English-only), an end-to-end ASR model, a conversational/chat agent, or creative paraphrasing—S1-mini is explicitly not a chat model and will only perform the trained normalization task.
Deployment and practical notes
- Recommended decoding: greedy (temperature 0) because the task is deterministic. Sampling introduces unwanted variance.
- Integrations: GGUF builds for llama.cpp / Ollama / LM Studio; transformers support (Qwen3 token type) available in recent versions of transformers. Always include the documented system prompt, the control line, and ensure the chat template's thinking block behavior matches training (enable_thinking=false).
- Licensing: Apache‑2.0 plus a naming clause that requires keeping the model name "S1-mini" by "Superwhisper" where used—check LICENSE before redistribution.
In short, S1-mini trades generality for a tight, predictable, and deployable normalization step that slots into any ASR → UI pipeline where readable, consistent text matters more than open-ended language generation.
