Recognizing spoken word can be improved when you have historical data related to a specific user (or a smaller group of users sharing the device).

Sure, anonymizing it and aggregating much larger numbers helps too. But modeling specific speech of one person, as well as the type of requests they make (to better understand/predict future recognition), can offer better improvements.