Last I looked, they used non-linear mixing in the analog circuity to down-convert ultrasonic audio into what looks like human speech to the devices.