Confucius4-R2T2: A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription
Daniel Huang, Qi He, Shaoxiong Duan, Yingjie Li, Qiao Cheng, Yitao Duan
NetEase Youdao · UC San Diego
Published November 4, 2026 · on the feed November 9, 2026 · ▲ 91 · arXiv 2611.06043
The abstract
Streaming speech recognition is increasingly consumed by machines, such as translators and voice agents, that act on each token as it appears. A token that is later revised costs wasted work or a wrong action. In this work, we present Confucius4-R2T2 (short for Real Real-Time Transcription), a stable-prefix streaming ASR model that never modifies its earlier output while maintaining high accuracy and low latency. R2T2 is obtained by post-training Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio while freezing its pre-trained audio encoder and projector. The adaptation is based on the concept of Longest Stable Prefix (LSP), a new notion that we introduce in this work. We prove that LSP is monotone under the prefix order, consistent with the full-audio output, and maximal among labels that equal a whole earlier output. Learning from LSP labels converts a full-context model into a stable-prefix streaming one without any architectural change. Extensive evaluations show that R2T2 brings streaming recognition close to the base model's full-context quality while delivering stable text with low emission lag. Its gains extend beyond the post-training languages, demonstrating a practical path from full-context ASR to stable streaming. We open-source the model weights and code at https://huggingface.co/netease-youdao/Confucius4-R2T2 and https://github.com/netease-youdao/Confucius4-R2T2.
The Brief
2:31A short podcast summary of the paper. Written and voiced by AI.
Transcript
Confucius4-R2T2. A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription. By Daniel Huang and five others, from NetEase Youdao and UC San Diego. Published on the fourth of November.
The output of streaming speech recognition is increasingly consumed by machines: translators and voice agents that act on each token the moment it appears. For a consumer like that, a token that is later revised is not a cosmetic flicker. It costs wasted work, or a wrong action that has already been taken. Accuracy and latency still matter, but a third requirement now sits beside them: what has been said should stay said.
Confucius4-R2T2, where R2T2 stands for Real Real-Time Transcription, is a streaming ASR model that never modifies its earlier output, while keeping accuracy high and latency low. The concept underneath it is new to this work: the Longest Stable Prefix, or LSP. The authors prove three properties of it. It is monotone under the prefix order. It is consistent with the output the model would give on the full audio. And it is maximal among labels that equal a whole earlier output.
Put in plainer terms, those three properties are what a streaming system would want from a commitment. Monotone means the committed text only ever grows: once something is in the prefix, it stays there. Consistent means that what was committed early agrees with what the model would have produced had it heard everything. And maximal means the prefix is not timid. Among labels of that kind, it commits as much as can be committed.
R2T2 is obtained by post-training Qwen3-ASR-1.7B on one thousand five hundred hours of Chinese and English audio, with the pre-trained audio encoder and projector frozen. The training labels are LSP labels. Learning from them is what converts a full-context model into a stable-prefix streaming one, and it does so without any architectural change. The model's structure stays as it was. What it is taught to emit is different.
Extensive evaluations show R2T2 bringing streaming recognition close to the base model's full-context quality, while delivering stable text with low emission lag. That is the trade the work was after: stability without giving up much of the quality a model gets from hearing the whole utterance. The gains also extend beyond the two post-training languages, Chinese and English. The model weights are released alongside the paper.
Here the contribution sits in the labels rather than the network. A new notion, with proved properties, defines what a stable prefix is. Post-training on labels built from it turns a full-context recogniser into one that never takes back what it has emitted, leaving the architecture untouched, and the authors present this as a practical path from full-context ASR to stable streaming.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
For a machine listening in, a rewritten word is not a flicker. It is a wrong action already taken.
You run a live transcriber whose text feeds a translator and a voice agent. Each token it emits is acted on the moment it appears.
The road ahead
- 1 · The problem ~7 min A live transcriber that rewrites words it already sent breaks the machines acting on them, and current fixes only make rewrites rarer or hide them.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Train the recognizer to emit only the longest prefix of its own full-audio answer that it will never revise, and say Wait for an unsettled rest.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Decode full audio, derive Longest Stable Prefix labels from progressive outputs, post-train the decoder on them, then stream text that never changes.
- 6 · The experiment ~6 min You’ll guess: After post-training to never revise its output, how did R2T2's recognition quality compare with its base model running on the complete audio?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.