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 ('Ale

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...