Fact · Speech recognition
70milliseconds (reduction in 90th percentile endpointer latency)
Sainath et al. report that a joint RNN-T endpointer model reached a word error rate of 6.8 percent and cut the 90th-percentile endpointer latency by 70 ms.
Checked against its source on October 17, 2026
In the source's own words
“obtains a WER of 6.8% and reduces EP50 and EP90 by 20ms and 70ms, respectively”
A Streaming On-Device End-to-End Model Surpassing Server-Side Conventional Model Quality and Latency, Tara N. Sainath et al., ICASSP 2020 (arXiv:2003.12710), 2020, Section 6.2.1.
Open the original- What it measures
- The endpointer is predicted inside the recognition model rather than by a separate voice activity detector.
- What it does not show
- The improvement is small, 70 ms, and it comes from one experiment on one system.