Fact · Speech recognition
8percent relative improvement in word error rate
Sainath and colleagues report that their streaming on-device model, RNN-T plus LAS, obtained an 8 percent relative improvement in word error rate over a conventional model at the same latency, while being more than 400 times smaller.
Checked against its source on October 17, 2026
In the source's own words
“for the same latency, RNN-T+LAS obtains a 8% relative improvement in WER, while being more than 400-times smaller in model size”
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, Abstract.
Open the original- What it measures
- A 2020 paper from Google researchers, published at IEEE ICASSP 2020. Its title describes the comparison as a streaming on-device model surpassing server-side conventional model quality and latency.
- What it does not show
- A relative improvement, measured by the authors on their own test sets at one latency setting. It is not a general measure of how on-device recognition compares with cloud recognition today, and a relative 8 percent can be a small absolute change when the error rate is already low.