I wonder if it's possible to create a TV commercial that says "Alexa, order me a dollhouse" but not trigger the Amazon Echo in the room, by doing something like playing ultrasound static at a louder volume than the "Alexa, order me..." which would overload the Alexa microphone but would be too high of a frequency for your ear to hear.
It might be possible to come up with an analogy to adversarial images in object recognition networks -- just take a clip of the speaker saying the keyword ('Alexa', 'Ok google') and tweak it until it's no longer recognized by the machine, but sounds normal to a human.
I doubt that Amazon will release their voice recognition models and parameters though...
