My guess, lots of user data + user corrections. When we correct what our phones mishear, that data helps train the speech recognition models to get more accurate over time.