Huang, Daniel, He, Qi, Duan, Shaoxiong, Li, Yingjie, Cheng, Qiao, Duan, Yitao · Nov 4, 2026
Confucius4-R2T2: A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription
Markdown · JSON · Original source
The authors of Confucius4-R2T2 present a streaming ASR model designed not to revise words it has already output. They argue this matters because translators and voice agents may act as each token appears: a later revision can waste work or trigger a wrong action. The model is made by post-training Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio, while freezing its audio encoder and projector. The authors introduce Longest Stable Prefix (LSP), a way to label output prefixes that remain stable as more audio arrives. They say learning from LSP labels makes a full-context model into a stable-prefix streaming model without changing its architecture. Their abstract reports streaming recognition close to the base model's full-context quality, with stable text and low emission lag, and says gains extend beyond the post-training languages. This suggests a route to speech recognition that machines can use as speech arrives, rather than waiting for the full audio; the paper's abstract does not give the measurements behind those reported results. The authors say they open-source the and code.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 4, 2026 (2026-11-04).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T01:40:24.576Z. First observed: 2026-11-05T05:06:41.791Z. Retrieved: 2026-11-10T01:39:45.025Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the abstract, which reports evaluation outcomes but gives no quantitative results or evaluation details to assess those outcomes.
Open question
What evaluation results quantify how close R2T2 comes to the base model's full-context quality and how low its emission lag is?
Source-stated
The authors present Confucius4-R2T2 as a streaming ASR model that does not revise earlier output while aiming for high accuracy and low latency.
a stable-prefix streaming ASR model that never modifies its earlier output while maintaining high accuracy and low latency.
Source-stated
They post-train Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio, freezing its pre-trained audio encoder and projector.
post-training Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio while freezing its pre-trained audio encoder and projector.
Source-stated
The authors say learning from Longest Stable Prefix labels enables streaming without an architectural change.
Learning from LSP labels converts a full-context model into a stable-prefix streaming one without any architectural change.
Source-stated
The abstract reports recognition close to the base model's full-context quality, low emission lag, and gains beyond the post-training languages.
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
Source-stated
Because machines may act on each token as it appears, revisions can waste work or lead to a wrong action; stable output is intended to address that problem.
A token that is later revised costs wasted work or a wrong action.
Source and contributor context
Huang, Daniel, He, Qi, Duan, Shaoxiong, Li, Yingjie, Cheng, Qiao, Duan, Yitao · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Pre-training
A model's initial training on a large dataset, before further adaptation for particular uses.
Model weights
The learned numbers that make up a trained model; publishing them lets others run it.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 18508383-94a7-4829-8147-e9d4854da441.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T01:40:24.576Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: c6a1759febe66614adb8a4fa491866ab5bd12c2bf5210b032f29941602fa22e9. Source text hash: 051d8ac507da1ae033195ab696194734420b223714bd5c32d41fc2fdb1bee884.
Revision and source-check history
Version 1 · 2026-11-10T01:40:24.576Z
Confucius4-R2T2: A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription
The authors of Confucius4-R2T2 present a streaming ASR model designed not to revise words it has already output. They argue this matters because translators and voice agents may act as each token appears: a later revision can waste work or trigger a wrong action. The model is made by post-training Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio, while freezing its audio encoder and projector. The authors introduce Longest Stable Prefix (LSP), a way to label output prefixes that remain stable as more audio arrives. They say learning from LSP labels makes a full-context model into a stable-prefix streaming model without changing its architecture. Their abstract reports streaming recognition close to the base model's full-context quality, with stable text and low emission lag, and says gains extend beyond the post-training languages. This suggests a route to speech recognition that machines can use as speech arrives, rather than waiting for the full audio; the paper's abstract does not give the measurements behind those reported results. The authors say they open-source the and code.
Original source
cs.AI · Article
Source details
Original source date: Nov 4, 2026 (2026-11-04).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T01:40:24.576Z. First observed: 2026-11-05T05:06:41.791Z. Retrieved: 2026-11-10T01:39:45.025Z.
Limits
The supplied evidence is the abstract, which reports evaluation outcomes but gives no quantitative results or evaluation details to assess those outcomes.
Open question
What evaluation results quantify how close R2T2 comes to the base model's full-context quality and how low its emission lag is?
Source-stated
The authors present Confucius4-R2T2 as a streaming ASR model that does not revise earlier output while aiming for high accuracy and low latency.
a stable-prefix streaming ASR model that never modifies its earlier output while maintaining high accuracy and low latency.
Source-stated
They post-train Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio, freezing its pre-trained audio encoder and projector.
post-training Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio while freezing its pre-trained audio encoder and projector.
Source-stated
The authors say learning from Longest Stable Prefix labels enables streaming without an architectural change.
Learning from LSP labels converts a full-context model into a stable-prefix streaming one without any architectural change.
Source-stated
The abstract reports recognition close to the base model's full-context quality, low emission lag, and gains beyond the post-training languages.
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
Source-stated
Because machines may act on each token as it appears, revisions can waste work or lead to a wrong action; stable output is intended to address that problem.
A token that is later revised costs wasted work or a wrong action.
Source and contributor context
Huang, Daniel, He, Qi, Duan, Shaoxiong, Li, Yingjie, Cheng, Qiao, Duan, Yitao · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Pre-training
A model's initial training on a large dataset, before further adaptation for particular uses.
Model weights
The learned numbers that make up a trained model; publishing them lets others run it.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 18508383-94a7-4829-8147-e9d4854da441.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T01:40:24.576Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: c6a1759febe66614adb8a4fa491866ab5bd12c2bf5210b032f29941602fa22e9. Source text hash: 051d8ac507da1ae033195ab696194734420b223714bd5c32d41fc2fdb1bee884.
passed · 2026-11-10T01:40:24.576Z · anthropic/claude-opus-5.5
Publication: 18508383-94a7-4829-8147-e9d4854da441
attribution — pass: The reading attributes the model, the LSP concept and the reported results to the paper's authors. Evaluation outcomes are framed as what the abstract reports, not as established facts. This matches the arXiv listing by Daniel Huang et al.
Authors:Daniel Huang, Qi He, Shaoxiong Duan, Yingjie Li, Qiao Cheng, Yitao Duan
summary — pass: The summary leads with the development: a streaming speech recognition model that does not revise output it has already produced. It gives the authors' motivation in their own terms (wasted work or a wrong action), the training method, the LSP idea, the reported results and the open release. Plain-language glosses (not revising already-output words; usable as speech arrives instead of after the full audio) make the core idea accessible. Remaining terms such as token and emission lag can be handled by tooltips. The question addresses a real gap in the abstract.
A token that is later revised costs wasted work or a wrong action.
date — pass: The originalPublished date, 2026-11-04, matches the v1 submission date on the arXiv page. It is not the later feed date of 2026-11-05.
[Submitted on 4 Nov 2026]
support — pass: Each claim is supported by abstract text. The closing significance sentence tracks the authors' own 'practical path' claim and is hedged with 'suggests'. It also notes that the abstract gives no measurements. No invented benefits, costs or numbers appear. The limitation statement accurately describes the evidence.
demonstrating a practical path from full-context ASR to stable streaming
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the original source. No better source media was found.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A model's initial training on a large dataset, before further adaptation for particular uses.
Definition referenceThe learned numbers that make up a trained model; publishing them lets others run it.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA model's initial training on a large dataset, before further adaptation for particular uses.
Definition referenceThe learned numbers that make up a trained model; publishing them lets others run it.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv