"S1-mini" by "Superwhisper" — LiteRT-LM conversion
This repository contains "S1-mini" by "Superwhisper" converted to the .litertlm format so it can run on-device with Google's LiteRT-LM runtime (Android, iOS, desktop).
- Original model: superwhisper/s1-mini (revision
88f6b15), a fine-tune of Qwen3-0.6B that cleans up speech-to-text transcripts. - This conversion is unofficial and is not affiliated with or endorsed by Superwhisper.
- It is used as the "Light" cleanup model in LocalType, an on-device voice keyboard for Android.
Files
| File | Description |
|---|---|
s1-mini_q8_ekv2048.litertlm |
Text model with dynamic int8 weights, 2048-token context, bundled tokenizer and chat template (648,634,192 bytes, SHA-256 d444ae7e52adab5f9eac36adb9609d1af7b76c73021ba5885f5b90d8503534fb) |
convert_s1.py |
The script used to produce it (litert-torch 0.9.4) |
LICENSE, NOTICE |
The original model's license and notice, unchanged |
Usage
Use the original model's exact system prompt and control line, with thinking disabled:
System: You are a text normalizer for speech-to-text transcripts. The input begins with a control line specifying the styling, structure, and context settings; clean the transcript to match those settings and output only the cleaned text.
User: [Styling: semi-formal] [Structure: lists] [Context: general]
<transcript>
See the original model card for the available settings.
Known limitations
In our tests the model occasionally changes spoken numbers or corrections (for example "ten thirty, actually no, make it eleven" became "11:30"). Review output where exact numbers matter.
License
Apache License 2.0 with an additional term, see LICENSE. Per that term, any use, distribution, or integration of this model must continue to identify it as "S1-mini" by "Superwhisper".
- Downloads last month
- 11
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support