Reproduce “Confucius4-R2T2: A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Daniel Huang, Qi He, Shaoxiong Duan, Yingjie Li, Qiao Cheng, Yitao Duan
R2T2 adapts Qwen3-ASR-1.7B for streaming speech recognition, training its decoder to release stable text while keeping the pretrained audio encoder and projector frozen. Its Longest Stable Prefix labels identify whole earlier outputs that match the model’s full-audio output. At 160-ms chunks and U=1, total macro error is 5.76%, versus 21.01% for the unadapted model; retrospective lag is not serving latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.