NetEase Youdao Open-Sources Streaming ASR Model Confucius4-R2T2

On September 17, NetEase Youdao open-sourced Confucius4-R2T2 (Real Real-Time Transcription), a streaming ASR model of roughly 2B parameters (some posts say 1.7B), built on Qwen3-ASR with a vLLM backend. Code and weights are live on GitHub, Hugging Face and ModelScope with an online demo, fully open and free to run. Multiple commentators argue it marks streaming ASR's shift from demo to infrastructure.

Confirmed

• Streaming decoding block size is configurable from 80ms to 2s, letting developers trade latency for accuracy; VraserX cautions that 160ms is the minimum streaming audio step, not end-to-end response latency, though some posts cite 200ms transcription latency.
• Core design is append-only: each word is committed once recognized and never rewritten, preventing unstable transcripts from polluting downstream agent state; Rohan Paul explains agents should consume speech incrementally but act only on committed text, with stable-prefix learning to prevent state pollution.
• A single Audio Encoder + LLM backbone serves both offline and streaming recognition via configurable chunk sizes, with no offline accuracy loss and no need to maintain two drifting models.
• The model is domain-aware, recognizing domain terminology.
• It supports an explicit wait state: per drcintas relaying the authors, the hard problem in streaming ASR is not predicting faster but judging whether a prediction is premature—the downstream cost of premature output exceeds that of a brief pause, so the model waits for evidence.

Why it matters

• Rohan Paul notes constantly rewritten transcripts are a serious production problem for voice agents; the stable-commit mechanism directly addresses it.
• Rohan Paul argues domain-aware streaming recognition, one stack covering all latency targets, and context as first-class input together separate infrastructure from demo; stable, domain-aware streaming recognition remains the real open question.
• VraserX holds that latency and responsiveness are nearly as important as model intelligence for voice AI—if the AI understands while you are still speaking, the whole interaction changes; he looks forward to what applications this open model enables.

2026-09-17 ~ 2026-09-18 · 21 related posts

Full story(2 episodes)→

• Episode 1: NetEase Youdao Open-Sources Streaming ASR Model Confucius4-R2T2(2026-09-17, 21 posts)
• Episode 2: NetEase Youdao Open-Sources Streaming ASR Model Confucius4-R2T2(2026-09-20, 3 posts)

• 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
• Voice AI's real battleground: latency may matter as much as intelligence — 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
• NetEase Youdao open-sources Confucius4-R2T2, a streaming ASR with 80ms chunks — rohanpaul_ai · 2026-09-17
• [source] NetEase Youdao open-sources Confucius R2T2, a 2B real-time speech model with tunable latency — dr_cintas · 2026-09-17
• NetEase Youdao open-sources R2T2, a 2B streaming ASR model with an explicit wait state — dr_cintas · 2026-09-17
• R2T2 reframes streaming ASR: the hard skill is knowing when to wait, not predicting faster — dr_cintas · 2026-09-17
• R2T2's design philosophy: knowing when a prediction is premature beats predicting faster — dr_cintas · 2026-09-17
• [source] R2T2: open-source streaming ASR model adds an explicit wait state instead of always guessing — dr_cintas · 2026-09-17
• NetEase Youdao open-sources Confucius4-R2T2, a 1.7B real-time streaming ASR model — _AndrewZhao · 2026-09-17
• NetEase Youdao open-sources Confucius4-R2T2, a 1.7B real-time streaming ASR model for voice agents — _akhaliq · 2026-09-18

6 near-duplicate retellings: dr_cintas · dr_cintas · socialwithaayan · dr_cintas · VraserX · rohanpaul_ai