Fact · Voice privacy and security
A 2018 study shows that a language model can be trained on client devices without exporting sensitive user data to servers, and that the approach matched or improved prediction recall.
Checked against its source on October 17, 2026
In the source's own words
“This work demonstrates the feasibility and benefit of training language models on client devices without exporting sensitive user data to servers.”
Federated Learning for Mobile Keyboard Prediction, Andrew Hard et al., arXiv, 2018, Abstract.
Open the original- What it measures
- The model was a next-word predictor for a phone keyboard. The training used federated averaging, which sends model updates from each device rather than the typed text.
- What it does not show
- It concerns text prediction, not speech recognition, and it is a design study. It does not show that any dictation product uses this method today.