The Confucius4-R2T2 system employs an append only transcription method that prevents committed text from being rewritten to ensure stability for downstream software. This automatic speech recognition tool allows developers to configure streaming chunks from 80 milliseconds to 2 seconds. The architecture supports local deployment for live captioning, speech translation, and voice agents.

Built on Qwen3-ASR, the 1.7 billion parameter model uses Longest Stable Prefix learning to decide when text is safe to emit based on audio context. Its LLM based decoder allows users to inject names, product terms, or industry jargon at runtime to steer recognition without modifying model weights. The system reduces transcription latency to 200 milliseconds to target failures where agents act on partial speech before a speaker finishes.