Most live transcription quietly edits itself: a word appears, then changes a moment later as more audio arrives. That is fine for captions people glance at, but a voice agent or a simultaneous translator that has already acted on the first version cannot take it back. NetEase Youdao's Confucius4-R2T2, described in a new preprint with open weights and code, is built so that text it has shown is never revised, only extended.
The team started from Qwen3-ASR-1.7B, which we review, and post-trained only its language decoder on about 1,500 hours of the company's own Chinese and English audio, leaving the roughly 300M-parameter audio encoder untouched. The trick is in the training labels. At each point in a clip, the target is the longest earlier output that still matches the start of the model's own full transcript, so the model learns from itself rather than from a separate teacher. A dedicated Wait token lets it hold off when the next word is not yet settled, and a reinforcement learning stage rewards accuracy while penalising repetition and invented continuations.
The results are strong for a model this small. With 160-millisecond chunks and one token held back for reconsideration, its average error across 13 English and Chinese test sets was 5.76%, against 5.15% for the base model working offline on complete audio. Run in the same tight streaming setting, the base model averaged 21.01%, so the post-training removes most of the penalty. R2T2 also beat the three commercial services in the comparison on average, including AssemblyAI at 8.27%, although two unnamed commercial systems were more accurate on Chinese. With 80-millisecond chunks, the paper reports average lag below half a second in both languages, at some cost in accuracy.
For anyone building on Qwen3-ASR, this is a practical upgrade path rather than a new architecture: the encoder and model structure are unchanged, so existing serving setups should carry over. The base model already offers streaming, but its partial text can change, which is the property R2T2 removes. Voice-agent builders who trigger actions on each recognised phrase are the clearest beneficiaries.
The limits deserve attention. This is a preprint and not peer reviewed, and the training audio is proprietary, so the gains cannot easily be reproduced. Lag was measured on the audio timeline only, without real processing, queueing or network time, and figures for the commercial services are approximate because their buffering is not disclosed. The paper does not state the licence for the weights, so check the repository before commercial use. Results beyond Chinese and English sit in an appendix, and testing on your own audio remains the sensible step before switching.