NetEase Youdao's Confucius4-R2T2 Streams Speech Without Ever Revising Committed Words

NetEase Youdao has released Confucius4-R2T2, a streaming speech-recognition model that permanently commits each emitted segment. Many streaming systems revise earlier words as more audio arrives, which causes visible caption changes and can send conflicting text to translation pipelines, voice agents, and other downstream software.

R2T2, short for Real Real-Time Transcription, builds on Qwen3-ASR and supports decoding chunks from 80 milliseconds to 2 seconds. The configurable chunk size lets developers balance responsiveness, compute overhead, and recognition accuracy while preserving append-only output.

How R2T2 commits text

R2T2 decides after each audio chunk which tokens have enough context to publish. Its training data combines stable prefixes, forced time alignment, and token-level audio segmentation. A Longest Stable Prefix learning method then teaches the model to emit confirmed text while retaining uncertain trailing tokens for later decoding.

The unfixed_token_num setting controls how many trailing tokens remain pending inside the decoder. Those tokens may change before publication, but text already returned to the application remains fixed. This gives subsequent decoding steps stable textual context and prevents clients from reconciling revisions.