The article mentions that the devices did manage to record conversations in at least one case. If so, they are clearly capable of recording conversations, just not optimized for it.
Also, how do you distinguish gunshots from background noise and triangulate the location of shots without first collecting raw audio from multiple devices and analyzing it?
Indeed, there's clearly some disinformation going on in the article:
"James G. Beldock, a vice president at ShotSpotter, said that the system was not intended to record anything except gunshots and that cases like New Bedford’s were extremely rare. “There are people who perceive that these sensors are triggered by conversations, but that is just patently not true,” he said. “They don’t turn on unless they hear a gunshot.” "
So apparently "the sensors", "They don’t turn on unless they hear a gunshot.". How, exactly, do they "hear a gunshot" if they're not (yet) turned on?
I suspect the truth is there's some software configuration that inhibits _recording_ of the sensor data until a gunshot-like event occurs (though if _I_ were designing this system there'd be at least a 30second or so buffer, so I could archive the sounds that if heard _before_ a gunshot as well as afterwards). But I'd hesitate trust that "configuration" to be particularly secure - much like the TSA "pornoscanners" - which in spite of claims of it being impossible, seem to be able to record images for the amusement of the operators and their friends…
They could be designed to run some local preprocessing and only record/forward audio when a gunshot seems significantly probable. (Not that I'd bet they actually do. We need public pressure to get them made that way.)
Also, how do you distinguish gunshots from background noise and triangulate the location of shots without first collecting raw audio from multiple devices and analyzing it?