Confucius4-R2T2 Learns Stable-Prefix Streaming ASR Without a Causal Encoder

This is a Plain English Papers summary of a research paper called Confucius4-R2T2 Learns Stable-Prefix Streaming ASR Without a Causal Encoder. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.

Stable-prefix ASR for transcripts that downstream systems can trust

Confucius4-R2T2 adapts Qwen3-ASR-1.7B with Longest Stable Prefix (LSP) learning so each released transcript is a prefix of every later one. The model is post-trained on about 1,500 hours of Chinese and English audio, while its pretrained audio encoder and projector remain frozen. Its goal is to make streaming text usable immediately without later revisions.

That guarantee targets a practical weakness in streaming ASR: an early word can be revised when more audio arrives, forcing a translator or voice agent to discard work or act on a mistaken transcript. R2T2 defines stability behaviorally: output at time t must be a prefix of output at any later time. Stability is separate from accuracy and latency, so a system that waits until the utterance ends would satisfy the definition but fail the real-time goal.

The model keeps Qwen3-ASR-1.7B’s bidirectional AuT encoder rather than making it causal, and supports configurable chunks from 80 milliseconds to 2 seconds. The same model can run in streaming or full-context mode, with no added layers or heads. The paper’s central claim is that stable output can be learned through decoder supervision without changing the architecture.

LSP labels turn full-context outputs into streaming supervision...

Continue reading the full paper summary on AIModels.fyi →