> This system detects door slams and popped volleyballs just as well as it detects gun fire (which is to say, imperfectly on all counts).
Sound classification moved far beyond what you’re claiming years ago.
If you set the threshold at “imperfectly” then no system will ever meet your bar, because perfect is an unreasonable threshold. However, the claim that we’re unable to differentiate between gunshots and car doors is extremely incorrect.
Yes there are false positives. This is an alert system and doesn't have to be perfect. This is the difference between a sensor in a self-driving car and a sensor for something like Shotspotter.
If a self-driving car sensor has a 1% fail rate, then it is useless and would cause crashes and death. It couldn't be used.
But if shotspotter is 1% inaccurate, then it just sends police to a street corner where there is actually just road construction or a volleyball popped. The consequences are pretty low to nil.
Many people might not know this all calls have priorities in a policing system, they are not treated equal. So Shotspotter alerts are very low priority to officers. The same is true for home alarm systems. Home alarm systems trigger false alarms constantly. Police treat them as very low priority as a result. So if a 911 call comes in or an eyewitness reported shooting, or even a wreckless driver or suspected DUI, the police will treat that first. In fact if they are on the way to a shotspotter alert and encounter something else, they will stop for it.
This is to say, that the false positives aren't as important as everyone makes them out to be. If a false positive happens, it means that a cop that's sitting on a random street corner gets moved to a street corner where the false alarm from Shotspotter fired. There's not really any harm being done here.
Having American cops rush to your door thinking violence is happening is not inconsequential. Just a few days ago they shot at a kid who was playing with fireworks, due to a false positive. Also, see swatting and its fatal consequences.
As much as I hate to participate in political discussions, I have to agree here. No US cop would ever respond to gunshots in a casual manner. Think about Swatting.
Sound classification under fairly well-controlled conditions could be very different from a real world streetcorner situation, where sound can bounce off surfaces and generate different signals depending on direction and distance of origin relative to the collecting microphone. High-quality omni-directional microphones can be expensive and I imagine there would be theft concerns if they used really good ones.
I recommend primo EM272 microphone capsules for use with https://github.com/hcfman/sbts-aru. They are high quality, very sensitive with high signal to noise ratio, lauded for nature recording use cases. They can be bought assembled for around 65 euros in the Netherlands. However these capsules are often found in much more expensive equipment.
Sound classification moved far beyond what you’re claiming years ago.
If you set the threshold at “imperfectly” then no system will ever meet your bar, because perfect is an unreasonable threshold. However, the claim that we’re unable to differentiate between gunshots and car doors is extremely incorrect.