Voice agents should consume speech incrementally but only act on committed text, because a fast transcript that mutates text can corrupt downstream agent state.

NetEase Youdao just open-sourced Confucius4-R2T2, a streaming ASR (Automatic Speech Recognition) model built exactly around that constraint.

It never rewrites committed text, i.e. my text is never gets rewritten underneath me.

That append-only behavior targets a very serioius production failure in voice agents, where software may act on partial speech before the speaker finishes.

Built on Qwen3-ASR, R2T2 uses Longest Stable Prefix learning to decide when text is safe to emit and when it needs more audio context.

The underrated detail in Confucius R2T2 is that because the decoder is LLM-based, context can be injected at runtime.

Names. Product terms. Industry jargon. Meeting topics. You steer recognition without touching the weights, which is a very different design choice from treating the acoustic model as a fixed black box.