Here's how it would work: An accelerometer samples a phone's vibration about 100 times per second, so it would be able to detect pairs of keystrokes, according to a Georgia Tech news release about this research. It would model "keyboard events" and determine where the pairs of keys are located on the keyboard, and how far apart they are. Then it would compare the results against a dictionary the researchers developed for this demonstration. The dictionary defines words based on their locations on a typical QWERTY keyboard, like left/right or near/far. So in Georgia Tech's example, the word "canoe" would translate to c-a, a-n, n-o, o-e possibilities. That works out to left-left-near, and so on. The location code is checked against the dictionary, and it turns up "canoe" as the most likely word.