Confucius4-R2T2 is a stable-prefix streaming automatic speech recognition model for truly real-time transcription.
Confucius4-R2T2 streams speech into append-only text via stable-prefix supervision, with public code for real-time voice agents and translation.
Key Takeaways
- NetEase Youdao released Confucius4-R2T2, a stable-prefix streaming ASR model for real-time voice agents and simultaneous translation, with public model, code, and demo.
- The model outputs append-only, stable text, reducing downstream complexity caused by repeatedly revised partial transcripts.
- Its core method, Longest Stable Prefix supervision, converts a full-context ASR model into a streaming one through post-training without changing the architecture.
- Built on Qwen3-ASR-1.7B, it claims near-offline accuracy with low emission latency and configurable chunk sizes to balance responsiveness and recognition.
- The release highlights cross-lingual potential for multilingual voice applications, though detailed supported-language and benchmark data are not provided.
Meta description: NetEase Youdao released Confucius4-R2T2, a stable-prefix streaming speech model for voice agents and translators, with append-only output and public code.

Featured image: Abstract friendly illustration style showing a stable stream of blocks flowing from a microphone into voice-agent and translator interfaces; no real people, no logos, no text in the artwork.
NetEase Youdao has released Confucius4-R2T2, a stable-prefix streaming automatic speech recognition model designed for real-time applications. It is aimed at builders of voice agents and simultaneous translators, who need transcript text that can be consumed immediately rather than repeatedly rewritten. The model is available at https://huggingface.co/netease-youdao/Confucius4-R2T2, with its implementation at https://github.com/netease-youdao/Confucius4-R2T2 and a live demo at https://r2t2.youdao.com/demo.
Append-only text reduces downstream friction
Streaming speech systems usually turn audio into partial text while a speaker continues talking. Conventional systems often revise those partial results, which can force downstream components to handle changing fragments before a sentence is complete. Confucius4-R2T2 instead produces append-only, stable text. In practical terms, once a stable prefix appears, the model does not go back and rewrite it, giving voice agents and translation systems a cleaner input stream.
Stable-prefix supervision changes the training target
The central method is Longest Stable Prefix supervision. Instead of training a system to produce the best possible transcript only at the end of an utterance, it teaches the model to identify the longest prefix of text that is already stable enough to output. That supervision lets NetEase Youdao transform a full-context speech recognition model into a streaming one through post-training, without architectural modifications.
The base model is Qwen3-ASR-1.7B. The announcement states that the model achieves near-offline recognition accuracy with low emission latency. It also supports configurable streaming chunk sizes, allowing builders to tune the balance between responsiveness and recognition behavior in their own systems.
What the release means for real-time voice products
For teams working on voice agents or simultaneous translation, the main value is integration simplicity. Stable output means a downstream application can consume text sooner, without waiting for a final revision or managing frequent corrections. The public model, code, and demo make it possible for developers to test that behavior directly with their own audio and use cases.
The announcement also describes promising cross-lingual generalization, which may make the model useful for multilingual voice experiences, though the source does not specify a full list of supported languages or formal benchmark scores. The project invites community feedback and public experimentation.
Tags: Confucius4-R2T2, streaming speech recognition, Qwen3-ASR-1.7B, voice agents, simultaneous translation
Comments (0)