NetEase Youdao released the Ziyue 4-R2T2 real-time ASR technical report on November 6 with 200,000 Hugging Face downloads
Confucius4-R2T2 topped the Hugging Face automatic speech recognition chart four days after release with 200,000 downloads across model and quantized versions, and its technical report shows a 5.76 percent macro-average error rate across 13 main test sets.
On November 6, NetEase Youdao's Confucius4-R2T2 had topped the Hugging Face automatic speech recognition chart just four days after open-sourcing, and cumulative global downloads of the model and its quantized versions reached 200,000, with 60,000 for the main model and roughly 140,000 for quantized versions combined. Community members have since produced quantized versions and adaptations for different runtime environments, and the model's technical report has been posted to a preprint server.
The question R2T2 addresses is not what was recognized, but which words can already be committed before a sentence has finished. The report handles this through longest stable prefix supervision, an explicit wait behavior and a streaming commit strategy, folding the decision of when to output into model learning and inference. R2T2 is built on Qwen3-ASR-1.7B and, without changing the architecture, learns streaming recognition and stable commit through post-training on about 1500 hours of Chinese and English audio.
Across all 13 main test sets, the quality-first configuration reached a macro-average error rate of 5.76 percent, below X-ASR at 8.97 percent and below the three commercial APIs evaluated in the report. Against a base model using the same 160 millisecond audio block and identical configuration, R2T2 cut the macro-average error rate from 21.01 percent to 5.76 percent. With 160 millisecond audio blocks and the low-latency configuration, average backtracking output delay was about 0.53 seconds for Chinese and 0.42 seconds for English.