R2T2 Streaming ASR Thread: 1.91% WER at 160ms Chunks, 0.42s English Lag
dr_cintas · x · 2026-11-06
Continuing the Confucius4-R2T2 paper walkthrough, drcintas details results at 160ms chunks across two configs: quality-first (U=1) hits 1.91% WER on LibriSpeech Clean, 8.81% on GigaSpeech Clean, 6.74% CER on WenetSpeech Meeting and 3.27% on CN-RealSI, with a 5.76% combined macro-average across 13 English and Chinese test sets. Latency-first (U=0) shows mean retrospective emission lag of 0.42s (English) and 0.53s (Chinese). Lag is measured on the audio timeline only, excluding computation, queueing, transport and rendering — so real-world user experience is likely even better.
More from Models
- scaling01: The cope for Chinese models is just distillation from Fable — scaling01 · 2026-11-07
- DeepSeek leads the accuracy-cost Pareto frontier "not even close", says casper_hansen_ — casper_hansen_ · 2026-11-07
- $50K live trading battles: AI bots currently edge out Claude and GPT — Mr_Hodlerr · 2026-11-07
- Brazil's Maritaca Releases S5-nano 4B, Claiming Best Open Weights in Brazil — ApprehensiveAd3629 · 2026-11-07
- Grok Bot's computer internet access issue is fixed — nima_owji · 2026-11-07
- Consistency test: Kimi K3.1 Pro passes 45% of VIALS tasks on all 3 tries vs 63% for GPT-6 Astra — ArtificialAnlys · 2026-11-07