NetEase Youdao open-sources R2T2, a 2B streaming ASR model with an explicit wait state
dr_cintas · x · 2026-09-17
NetEase Youdao launched Confucius R2T2, a 2B-parameter real-time streaming ASR model that is fully open source and free to download.
- Decoding chunks configurable from 80ms to 2s, letting you set your own latency-accuracy tradeoff
- Key design insight: most streaming ASR systems must emit something every step, falling back on language priors when audio is ambiguous; R2T2 introduces an explicit wait state — it holds when acoustic evidence is insufficient and commits when it arrives
- Claims near-offline accuracy in real time, with English and Mandarin demos
- Code, weights (GitHub/HF/ModelScope) and a hosted demo are live, with a vLLM backend; full technical report to come
More from coding & agent
- Open weights aren't enough: four ownership tests for your personal AI memory and harness — sujingshen · 2026-09-17
- Asimov goes live every Friday to demo its AI coding agent in action — freelerobot · 2026-09-17
- Diorama gives OpenAI Codex coding agents a visual office you can watch work in real time — davidfromkansas · 2026-09-17
- Code-first, UI on top: building bespoke brand design tools with AI — floguo · 2026-09-17
- Study of 7 models across Claude Code, Codex, Pi: harness barely affects success but swings cost — DavideCrapis · 2026-09-17
- AI trading bot built with Jev is down 85%, owner shrugs it off — generativist · 2026-09-17