They make it a wearable to counteract the nulls caused by the multiple speakers. Random movement causes the nulls to not stay in one place for very long, limiting how much of a word or sentence a microphone could pickup. You could build a phased array of these that could quickly move a hotspot of ultrasonic noise all over the room. You could them position on the ceiling with a fixed radius between them to make sure that the highest pressure occurs at about waist level, where phones in pockets, smart watches on wrists, and smart speakers on tables would reside. Another idea would be a ball of these in the center of the room and have it move up and down to get the best average coverage around the room.
Smart assistants are usually not recording really high quality audio, it takes more time to process it and more time to send it back home so they are going to a lower sample rate than typical voice recorder app would use. Siri uses a 16KHz sample rate (Fs=16KHz) which is enough to put the whole human vocal range in the 1st Nyquist zone (less than Fs/2). Playing a sound at 26KHz (3rd Nyquist zone, >Fs but <1.5*Fs) is going to cause a reflection across Fs. So the 26KHz tone, sampled at 16KHz, creates a tone at 10KHz which could be enough to confuse a naive implementation of a smart assistant. Ideally, you want fix this by either installing an analog filter so the ultrasonic noise can never reach the ADC or sample the whole range (up to 44.1KHz is a good start) and filter digitally.
There is a paper called DolphinAttack [1] where they attempted to use the ultrasonic audio band as an inaudible attack vector. You could play an ultrasonic noise that no one can hear except for the smart assistant.
Smart assistants are usually not recording really high quality audio, it takes more time to process it and more time to send it back home so they are going to a lower sample rate than typical voice recorder app would use. Siri uses a 16KHz sample rate (Fs=16KHz) which is enough to put the whole human vocal range in the 1st Nyquist zone (less than Fs/2). Playing a sound at 26KHz (3rd Nyquist zone, >Fs but <1.5*Fs) is going to cause a reflection across Fs. So the 26KHz tone, sampled at 16KHz, creates a tone at 10KHz which could be enough to confuse a naive implementation of a smart assistant. Ideally, you want fix this by either installing an analog filter so the ultrasonic noise can never reach the ADC or sample the whole range (up to 44.1KHz is a good start) and filter digitally.
There is a paper called DolphinAttack [1] where they attempted to use the ultrasonic audio band as an inaudible attack vector. You could play an ultrasonic noise that no one can hear except for the smart assistant.
[1] https://gangw.cs.illinois.edu/class/cs598/papers/ccs17-hidde...