NetEase Youdao open-sources Confucius4-R2T2 streaming ASR model
Summary
NetEase Youdao has released Confucius4-R2T2, an open-source streaming Automatic Speech Recognition (ASR) model designed to enhance the reliability of voice agents by adopting an append-only output mode. This feature ensures that once a transcript segment is committed, it cannot be altered, preventing the common failures in voice assistants that arise from acting on incomplete or changing text. Built upon the Qwen3-ASR model, R2T2 enables high-accuracy recognition for both streaming and offline use without degrading performance, highlighting its unified architecture that eliminates the need for separate models.
Analysis
Qwen3-ASR: Qwen3-ASR is an LLM-based automatic speech recognition model from the Alibaba Qwen team that serves as the architectural foundation for Confucius4-R2T2. Its decoder allows runtime injection of context such as domain terms or hotwords, which Confucius4-R2T2 leverages for customizable recognition without retraining. NetEase Youdao: NetEase Youdao is a Chinese technology company focused on online education, translation, and AI-driven language tools. It recently open-sourced Confucius4-R2T2, a streaming ASR model designed to support production voice agents by ensuring stable, append-only transcript output. Confucius4-R2T2: Confucius4-R2T2 is a low-latency streaming Automatic Speech Recognition model developed by NetEase Youdao and released as open source in 2026. It is built on the Qwen3-ASR foundation with Longest Stable Prefix learning to emit committed text without revisions, enabling reliable integration with downstream LLM agents and real-time applications. Open Source Release: Confucius4-R2T2 was released as open source by NetEase Youdao with inference code, a vLLM backend, and support for both streaming and offline modes under a dual licensing model. Unified Architecture: Confucius4-R2T2 uses a single Audio Encoder plus LLM foundation that supports configurable chunk sizes for streaming while preserving full offline accuracy without maintaining separate models. Voice Agent Reliability: The append-only design of Confucius4-R2T2 addresses production failures in voice agents where partial or mutating transcripts can corrupt agent state before speech is complete.
Categories
techcryptomachine_learningaiai_agents