Confucius4-R2T2 Records 5.76% Error Rate at 160 Milliseconds

Confucius4-R2T2, a streaming speech-recognition model trained on Chinese and English audio, recorded a 5.76% error rate at 160 milliseconds, targeting captions that update as people speak.

In the reported test, the model used one provisional token. Its base model, Qwen3-ASR-1.7B, had a 21.01% recognition error rate under the same conditions. The comparison measures performance at the same latency setting, making the two figures directly comparable.

The system is designed to keep text from changing after it appears. Its Longest Stable Prefix method trains the transcript to advance in step with the portion of a full-audio transcript that remains consistent as more speech arrives. That matters in live settings, where revisions can disrupt captions or software processing the words in real time.

Confucius4-R2T2 also has a “Wait” action. It can ask for more audio before committing to a transcript, rather than producing text from an incomplete segment. The model adapts Qwen3-ASR-1.7B through post-training on 1,500 hours of Chinese and English audio, using the Longest Stable Prefix method to support streaming output without changing the model architecture.

Potential uses include live captioning, simultaneous speech translation and systems that pass recognized speech into later language-processing steps. The model’s focus is stable output during streaming. By comparison, Meta’s Omnilingual Automatic Speech Recognition system supports more than 1,600 languages.

The model’s weights and code are publicly available.