You can also trigger wake words like "Ok Google" or "Alexa" by using _harmonics_ of normal human voices that are outside the range of audible sound. The key is that the mic can't differentiate between the harmonics and actual speech of normal human pitch and so the trigger is set off, but the sound isn't audible to humans.
I don't have a link to the paper handy (sorry!) but IIRC I found a white paper on ArXiv called "Dolphin Attack" or similar that demonstrated this. It was a fun read.
While technically that sounds like a cool hack, it sounds like a dumb thing to actually allow to happen. "Okay Google" or "Hey Alexa" outside of the frequencies reproducible by human voices should be filtered out completely. At least, from my comfy chair nit picking someone else's work. Of course, by not filtering the acceptable bands allows them to do all of the ad tracking/rights management things they are allowing to occur. The fact that Alexa/Google is able to do these ad tracking/rights management is just another example showing that the mic is listening 100% of the time.
I don't have a link to the paper handy (sorry!) but IIRC I found a white paper on ArXiv called "Dolphin Attack" or similar that demonstrated this. It was a fun read.