Ultrasonic waves and machine learning are being brought together by Project Telepathy, a new research project out of the University of Bristol. The project uses machine learning to figure out what a user is saying silently via the movements of their facial muscles. From there, a very narrow ultrasonic beam from a specialized speaker mounted on the subject’s forehead or chest is emitted to transfer the message. The speaker generates the words that the sensors think a subject is mouthing and convey them via an ultrasonic beam that only those in its path can hear by using an identical speaker.
The solution relies on an electromyographic system mounted on a subject’s face. Four electrodes read their facial muscle movements in detail to figure out what words they’re mouthing. It’s not a wild guess to venture that the system is limited for now; current accuracy is at approximately 80 percent but that’s only across ten words that the platform can currently recognize. A machine learning model was trained through repeated exposure to both mouthed and spoken words over time, and with proficiency in discerning spoken words came proficiency in discerning mouthed words. Part of the training consisted of a study involving two female and four male participants, who put on the electrodes, then spoke a set of words as instructed. Those words were left, right, up, back, stop, turn, yes, no, faster, and forwards. The sessions each lasted 40 minutes and saw participants changing up the pronunciation and the word spoken every five attempts. Participants also alternated between speaking and mouthing the test words.
The system uses one of two different speaker relays and each works in a somewhat different manner but achieves the same goal. They both have a large number of small speaker nodes bundled closely together. A forehead-mounted system points where the user’s head points, and uses eye tracking for more precise aim. This allows a user to simply look at somebody to speak only to them, whether they’re level with their position, above, or below, somewhat akin to targeting non-player characters to talking to them in some classic video games. The chest-mounted system is somewhat more precise because the array has a larger area to work with, and uses a camera and pointer stick for targeting. In either case, the system “paints” sound in a 3D spectrum, then sends those sound waves traveling in a specific direction with no leakage, allowing them to hit only the targets in front of them. It remains to be seen whether this solution will ever be commercialized but an update on its progress may follow in the future.