Smart assistants like Siri have already been the target of jokes when it comes to misinterpreting their activation hotwords. However, users have come to realize how costly these errors can be, especially with regard to possible privacy issues and embarrassing situations. The companies behind these smart assistants are always trying to improve the way they detect hotwords properly, and Google is now trying to virtually crowdsource the data that Google Assistant needs to learn, but without going through the potentially confusing and controversial route from the cloud.
Google Assistant, like many of Google’s services, is able to improve its services and algorithms through the data it collects from users. With the Assistant, however, Google took great care not to send unnecessary voice data to its cloud, especially after the big scandals that rocked the AI assistants market. Technically, Google says that it only starts recording voice clips after being triggered by the hotword “Hey, Google”.
Unfortunately, it is precisely this hotword that is hindering Google Assistant and needs more data about error triggers to improve its recognition capabilities. To this end, Google is applying a federated learning strategy that uses the phrases spoken by Google Assistant users in a more privacy-friendly way.
In short, raw voice data is not sent to Google’s servers and is instead processed locally. Only the anonymous data model is sent to Google and combined with data from other users for the purpose of improving AI algorithms and processes. Voice recordings are encrypted and stored on the device and deleted after about two months when they are no longer needed.
To be clear, the feature is disabled by default and you will need to enable it if you want to help improve the performance of the Google Assistant. Google is applying similar federated learning strategies to its other applications and services. He also started to outsource certain Android application data to improve download and installation speeds from the Google Play Store.