Append-only streaming ASR: why voice agents should only act on committed text

rohanpaul_ai · x · 2026-09-17

Rohan Paul breaks down NetEase Youdao's open-source Confucius4-R2T2 (built on Qwen3-ASR): voice agents should consume speech incrementally but act only on committed text, since mutable transcripts corrupt downstream agent state. R2T2's append-only design uses Longest Stable Prefix learning to decide when text is safe to emit. Underrated detail: the LLM-based decoder allows runtime injection of names, product terms, and jargon without touching weights — a stark contrast to treating the acoustic model as a fixed black box.

Related event: NetEase Youdao Open-Sources Streaming ASR Model Confucius4-R2T2(8 posts)→

Original post →

More from coding & agent

coding & agent channel →