Confucius4-R2T2: A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription
This paper presents Confucius4-R2T2, a stable-prefix streaming automatic speech recognition (ASR) model designed for real-time applications such as voice agents and simultaneous translators. Traditional streaming ASR systems often revise partial transcripts as more context arrives, which wastes compute and causes errors in downstream agents. To solve this, the authors introduce the Longest Stable Prefix (LSP) concept, proving it to be monotone, consistent with full-audio outputs, and maximal. By post-training Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio using LSP supervision while freezing the audio encoder and projector, R2T2 produces append-only, stable text without architectural modifications. Extensive evaluations demonstrate that R2T2 approaches base-model full-context quality with low emission lag and strong cross-lingual generalization. The code and model weights are open-sourced.