Meet S1-mini: Superwhisper's 462 MB Open-Weights Text Normalizer That Turns Raw ASR Transcripts Into Clean Written Text
Superwhisper released S1-mini, a 0.6B open-weights text normalizer fine-tuned from Qwen3-0.6B that converts raw ASR transcripts into clean written text The model achieves 94.8% token accuracy on a held-out set of 7,519 cases and runs on a laptop CPU as a 462 MB Q4_K_M GGUF file S1-mini is controlled via a three-axis system (Styling, Structure, Context) with independent values for casual/formal tone, prose/list format, and general/email context Unlike S1-Voice and S1-Language (cloud-only services
Analysis
TL;DR
- Superwhisper released S1-mini, a 0.6B open-weights text normalizer fine-tuned from Qwen3-0.6B that converts raw ASR transcripts into clean written text
- The model achieves 94.8% token accuracy on a held-out set of 7,519 cases and runs on a laptop CPU as a 462 MB Q4_K_M GGUF file
- S1-mini is controlled via a three-axis system (Styling, Structure, Context) with independent values for casual/formal tone, prose/list format, and general/email context
- Unlike S1-Voice and S1-Language (cloud-only services), S1-mini is fully self-hostable under Apache 2.0 with a naming clause
- The model is constrained by design: it removes fillers, resolves self-corrections, applies punctuation, and normalizes spoken numbers/dates/emails without adding or altering factual content
Why It Matters
S1-mini fills a critical gap in the speech-to-text pipeline by offering a lightweight, deployable solution for post-ASR text normalization that can run entirely on-premise or on-device. For AI practitioners building voice-driven applications, this means they can transform raw ASR output into polished, human-readable text without relying on cloud APIs, addressing data privacy concerns in regulated industries like healthcare and legal.
Technical Details
- Architecture: Fine-tuned from Qwen/Qwen3-0.6B with 596M unique parameters (0.44B non-embedding), 28 layers, 16 query heads and 8 key/value heads using GQA, and BF16 weights
- Control Interface: Three-axis control line placed above the transcript: Styling (casual/semi-casual/semi-formal/formal), Structure (prose/lists), and Context (general/email)
- Performance: 94.8% token accuracy on 7,519 held-out cases, 11.6% text-edit error rate, 99.3% greeting line identification, 97.9% sign-off detection, 92% exact email reproduction
- Deployment: Q4_K_M GGUF build is 462 MB, runs on laptop CPU, Apache 2.0 licensed with naming clause; English-only in v1 with ~1,000 token input recommendation
- Critical Implementation Requirements: Must set
enable_thinking=Falseand decode greedily (temperature=0); the Qwen3 chat template defaults to thinking mode which breaks output if not disabled
Industry Insight
- Privacy-First Voice Pipelines: Organizations handling sensitive audio (healthcare dictation, legal proceedings, financial calls) can now deploy text normalization locally, keeping transcripts within their VPC while still achieving professional-quality output
- Cost Optimization: The 462 MB model eliminates per-token API costs for text normalization, making it economically viable for high-volume applications like live captioning or voice-to-CRM entry
- Integration Caution: Developers must carefully handle the
enable_thinking=Falserequirement and greedy decoding; the mismatch between the app's five-stop tone slider and the open weights' four trained values suggests potential for user confusion that could affect adoption
Disclaimer: The above content is generated by AI and is for reference only.