Fact · Speech recognition
450milliseconds (median endpointer latency)
Sainath et al. report that with an acoustic endpointer alone, their on-device model reached a median endpointer latency of 450 ms, and its word error rate rose from 6.2 to 7.4 percent.
Checked against its source on October 17, 2026
In the source's own words
“With the acoustic endpointer alone, we degrade the WER from 6.2% (no EP) to 7.4% to achieve a 450ms EP50 latency and 860ms EP90 latency.”
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, Table 4.
Open the original- What it measures
- Experiments on Google's on-device RNN-T plus LAS system, on the paper's internal test set. The endpointer latency is the time from the end of the user's speech to the microphone closing.
- What it does not show
- A single system on one test set. The 450 ms figure is the best operating point the authors chose by sweeping endpointer settings, so it is a trade-off, not a fixed property of streaming.