s1-mini: Ultra-lightweight ASR post-processing model based on Qwen3-0.6B

This is a small language model responsible for the latter part of the speech recognition (ASR) pipeline, specifically text normalization and post-processing. It is fine-tuned from Qwen3-0.6B and optimized to a scale of approximately 750 million parameters. This approach integrates sentence structure organization, which was previously handled by separate modules in existing ASR systems, into a single LLM.

The core functions include text normalization, inverse text normalization, punctuation insertion, and Truecasing. It enhances readability by adding grammatical completeness to raw text extracted from speech. It also supports post-processing that considers conversational context, enabling the generation of natural sentences.

It is specialized for the text generation and refinement stages of the overall ASR pipeline. With its lightweight structure rather than a large-scale model, it facilitates real-time processing even in low-spec environments. It is released as part of the Superwhisper project and allows for immediate inference via the Hugging Face Transformers library.