NetEase Youdao Open-Sources Confucius4-R2T2 Streaming ASR Model

On September 17, NetEase Youdao open-sourced the streaming speech recognition model Confucius4-R2T2 (Real Real-Time Transcription) on GitHub. Built on Qwen3-ASR, it targets true low-latency, high-accuracy streaming recognition and offers an append-only, no-rewriting solution to the state-pollution problem in production voice agent environments. Several bloggers see this as marking streaming ASR's shift from demo to infrastructure.

Confirmed

• The model is built on Qwen3-ASR with configurable decoding block granularity starting at 80ms; 160ms is the minimum streaming audio step, not end-to-end response latency—a conceptual distinction blogger VraserX specifically highlighted.
• The core design is append-only: committed text is never rewritten, preventing fluctuating streaming transcripts from polluting downstream agent state; blogger Rohan Paul explains the idea is to let the agent consume speech incrementally, act only on committed text, and use stable-prefix learning to avoid state pollution.
• Architecturally, a single Audio Encoder + LLM backbone serves both offline and streaming recognition, tunable between latency and quality via configurable chunk sizes, with no accuracy loss on the offline side—no need to maintain two models that would drift apart.
• It is domain-aware and can recognize domain terminology.

Why it matters

• Rohan Paul notes that constantly rewritten transcripts are a serious production problem for voice agents, corrupting downstream agent state; R2T2's stable-commit mechanism hits this pain point directly.
• Rohan Paul concludes that domain-aware streaming recognition, a single stack covering all latency targets, and treating context as a first-class input rather than an afterthought—together these mark the line between "infrastructure" and "demo"; stable, domain-aware streaming recognition remains the real open question.

2026-09-17 ~ 2026-09-17 · 7 related posts

• [source] NetEase Youdao's Open-Source Confucius R2T2: 160ms Is Streaming Step, Not Latency — VraserX · 2026-09-17
• NetEase Youdao open-sources Confucius R2T2: 160ms streaming slices for speech AI — VraserX · 2026-09-17
• Append-only streaming ASR: why voice agents should only act on committed text — rohanpaul_ai · 2026-09-17
• [source] NetEase Youdao open-sources Confucius4-R2T2, append-only streaming ASR built on Qwen3 — rohanpaul_ai · 2026-09-17
• One audio encoder + LLM serving offline and streaming ASR without accuracy tax — rohanpaul_ai · 2026-09-17
• Domain-aware streaming recognition is what separates infrastructure from a demo — rohanpaul_ai · 2026-09-17
• [source] NetEase Youdao open-sources Confucius4-R2T2, a streaming ASR with 80ms chunks — rohanpaul_ai · 2026-09-17