Confucius4-R2T2: A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription

Brief overview

Confucius4-R2T2 is a stable prefix streaming speech recognition model for real-time applications, which converts a full context ASR model into a streaming model through post-training without architecture modification. Built based on Qwen3-ASR-1.7B, R2T2 achieves near-offline recognition accuracy and low transmission delay, supports configurable streaming block sizes, and demonstrates promising cross-lingual generalization capabilities. This model and implementation have been publicly released.

Article

Excited to share Confucius4-R2T2, our stable-prefix streaming ASR model designed for real-time applications!

Unlike conventional streaming ASR systems that repeatedly revise partial transcripts, R2T2 produces append-only, stable text that downstream systems such as voice agents and simultaneous translators can consume immediately.

The key idea is Longest Stable Prefix (LSP) supervision, which enables us to transform a full-context ASR model into a streaming one through post-training, without architectural modifications.

Built on Qwen3-ASR-1.7B, R2T2 achieves near-offline recognition accuracy with low emission latency, supports configurable streaming chunk sizes, and demonstrates promising cross-lingual generalization.

We've made the model and implementation publicly available:

🤗 Model: https://huggingface.co/netease-youdao/Confucius4-R2T2

💻 GitHub: https://github.com/netease-youdao/Confucius4-R2T2

🎙️ Live demo: https://r2t2.youdao.com/demo

We'd love to hear your feedback and see what the community builds with it!