NetEase Youdao Open-Sources Confucius4-R2T2: Streaming ASR That Never Rewrites Output

NetEase Youdao published a paper on arXiv on November 6 and open-sourced Confucius4-R2T2, a streaming speech recognition model that "never rewrites." As drcintas noted in a series of deep-dive posts, a persistent pain point of streaming ASR is that models keep revising text they have already output, causing subtitles to flicker and voice agents to repeatedly reprocess the same sentence—an issue that has evolved from a UX problem into a systemic one. R2T2's promise is that once text is emitted, it is never modified, which is especially valuable for machine scenarios like translators and voice agents that consume recognition results token by token, avoiding wasted work from revisions.

Confirmed

• The model is built on Qwen3-ASR with no architecture changes; the audio encoder and projector stay frozen throughout, with only the language decoder updated.
• Post-training used roughly 1500 hours of Chinese and English audio in a two-stage pipeline.
• The core contribution is not a new architecture but a new training objective, LSP (Longest Stable Prefix): at each timestep, it identifies the longest text prefix already "committed" by a full-context model as the training target, using an explicit mechanism to guarantee "never rewrite."
• At 160ms chunks, the quality-first configuration (U=1) achieves WER 1.91% on LibriSpeech Clean, with results on GigaSpeech Clean as well (the paper notes differences between two configurations).
• Streaming gains extend to 18 test sets across 10 languages.

Unconfirmed

• drcintas points out the cross-lingual transfer mechanism is unclear: training used only Chinese and English, yet streaming gains generalize to many languages, with no explanation in the paper.
• The chunking strategy uses fixed-length chunks and is not adaptive; the paper does not discuss this in depth.

Why it matters

• For downstream systems that consume ASR output token by token (translators, voice agents, subtitles), "never rewrite" eliminates reprocessing and flickering—a feature with clear engineering value.
• LSP, as a formalized training objective, turns the behavioral constraint of a "stable prefix" into a learnable goal, offering a transferable idea for streaming generation models.

2026-11-06 ~ 2026-11-06 · 8 related posts

• [source] NetEase YouDao open-sources Confucius4-R2T2, a stable-prefix streaming ASR that never revises output — _akhaliq · 2026-11-06
• Streaming ASR's Hidden Problem: Models Keep Rewriting Output — R2T2's Fix — dr_cintas · 2026-11-06
• R2T2's Real Contribution Is a Learning Target: Formalizing Longest Stable Prefix — dr_cintas · 2026-11-06
• How LSP Training Makes R2T2 Never Rewrite: Mechanism-to-Behavior Mapping — dr_cintas · 2026-11-06
• R2T2 Streaming ASR Thread: 1.91% WER at 160ms Chunks, 0.42s English Lag — dr_cintas · 2026-11-06
• R2T2 critique: unexplained cross-lingual transfer and fixed chunk sizes cap streaming ASR — dr_cintas · 2026-11-06
• [source] R2T2 streaming ASR: frozen audio encoder, decoder-only tuning with SFT then GRPO — dr_cintas · 2026-11-06

1 near-duplicate retellings: dr_cintas