160ms·U=1 average error rate 5.76%…Korean FLEURS 7.52%

NetEase Youdao researchers released Confucius4-R2T2, a streaming speech recognition model that emits only finalized transcript segments and never revises text once output. At a 160ms chunk size with U=1 (one temporary token held back), the model averaged a 5.76% error rate across 13 English and Chinese evaluation sets, using word error rate for English and character error rate for Chinese. On Korean FLEURS the error rate was 7.52%. The design matters because immutable partial output suits real-time applications such as live captioning and voice agents, where rewriting already-displayed text is costly. Next, watch for peer comparison of latency-accuracy tradeoffs at other chunk sizes and U values, plus independent replication on Korean and other non-English benchmarks.