NetEase YouDao open-sources Confucius4-R2T2, a stable-prefix streaming ASR that never revises output
_akhaliq · x · 2026-11-06
NetEase YouDao released Confucius4-R2T2 (Real Real-Time Transcription), a streaming ASR model that never modifies its earlier output — critical for machine consumers like translators and voice agents that act on each token.
- Built on a new concept, Longest Stable Prefix (LSP), proven monotone under the prefix order and consistent with full-audio output
- Post-training of Qwen3-ASR-1.7B on 1,500 hours of Chinese/English audio, freezing the pre-trained audio encoder and projector — no architectural changes needed
- Eval shows streaming quality close to the base model's full-context accuracy with low emission lag; gains generalize beyond the post-training languages
- Model weights and code are open-sourced
More from Models
- OpenAI researcher: 800M-model intuitions are useless for GPT-6-scale training — willdepue · 2026-11-06
- On the METR graph: insiders seeing AECI tasks is weak evidence, not zero evidence — gleech · 2026-11-06
- Daily AI brief: GPT-6.2 imminent, Claude Personal Finance beta, METR says Anthropic progress up 1.9x — testingcatalog · 2026-11-06
- METR report researcher quotes lean bullish: Opus 4.5 coding plus faster experiment cycles — gleech · 2026-11-06
- gleech: dropping the three weak benchmarks raises the estimated slope, but the y-axis is meaningless — gleech · 2026-11-06
- Epoch AI review: OSWorld tasks are simple, unstable, and barely test GUI use — gleech · 2026-11-06