To deploy this in a real world scenario seems implausible to me. You would need specific models for every person, laptop and router combination, as the physiology of the hands, the gestures when typing, and the location and power of the laptop's and router's antennas can all be different.

The only relevant attack could be to train on a specific target (person+laptop+location) for a given amount of time to achieve high accuracy, but to do this you already have to know what the person is typing...