This does make use of multiple microphones and timing between sound waves arriving. To my understanding, the machine learning part allows it to work even as the speaker/wearer move relative to each other, such as if you turn your head.