The device's current version is still a prototype, consisting of a speaker worn on the forehead or chest and electrodes placed on the lips and jaw. Those pick up electrical signals made by muscles in the face while a person talks, meaning someone wouldn't have to actually speak for the device to translate motion into words. The researchers used a machine learning algorithm that recognizes which muscular signals produce certain words. Those are emitted by wearable speakers at ultrasonic frequencies in a beam so narrow only those in its path would hear it.
The system is still very much in development: it can only recognize ten basic words and only got them right 80 percent of the time. But it could be useful for anyone in the market to speak securely over distance, like spies or soldiers in the field. It's a neat development in the field of nonverbal communication, the kind of research akin to Disney's experiments a few years ago sending messages through electric current.