It wouldn't be difficult to detect false positives based on the signature of a gun shot vs other loud noises.

I was gonna say, having heard a gunshot as well as the other stated things that could sound like gunshots, I think gunshots are categorically louder and more identifiable. That said, I'm not a gun expert so I could be wrong for different types of guns, etc.

Lots of differences that can be pulled out algorithmically. Gunshots are much sharper, as they're mechanical/aerodynamic rather than chemical, and are linear rather than point sources. Receivers placed across the city will record differently dopplered shockwaves based on how close they are to being in line with the bullet's trajectory. If you really wanted to, I bet that you could go around the city after installation and set off firecrackers in a grid to map out the city's acoustics, then reconstruct bullet trajectory down to thirty or forty meters. Which'd be pretty cool.

Of course, all that goes away in the case of subsonic rounds. Then your only real difference is that a gun propellant is going to have a much faster velocity than the gunpowder in fireworks and the explosion is going to be rather better contained. You'll probably still be able to differentiate, but I'm not sure how effectively, especially over long distances.

> Of course, all that goes away in the case of subsonic rounds.

I'm pretty sure that presents a problem. Many handgun rounds have subsonic muzzle velocities, and most are subsonic within 50 feet. And according to the FBI's Uniform Crime Reports[1], almost half of all murders are committed with handguns. Rifles are used in around 2-3%. (For comparison, 5-6% of murders are committed with bare hands.) Those numbers underestimate the ratio of rounds fired by handguns vs rifles, as being shot with a rifle is far more likely to result in death than being shot with a handgun. If the victim doesn't die, they won't show up in those FBI stats.

1. FBI Uniform Crime Report on murder weapon type: https://www.fbi.gov/about-us/cjis/ucr/crime-in-the-u.s/2013/...

I figured the numbers were something like that. Darn.

I bet if you were given six months and high quality recordings of different types of gunfire you could figure out how to do it algorithmically.

It's all about having a large enough corpus of input, both of the type you're looking for and the type you're looking to exclude. This is an undergraduate level problem to solve.

I would be very surprised if you could figure out a way to reliably discriminate between the sound of a gunshot and the sound of a car backfiring given that they're basically the same thing (that is, the sound of a gas rapidly expanding in a metal tube).

Yeah, well you'd be wrong. Anyone who's spent time around guns would instantly recognize the difference, and training a computer to do the same thing isn't even hard, it just takes a lot of patient work.

The most important difference here is a gun is pushing a projectile out and a car isn't.

Differentiating between cap guns, starter pistols and fireworks might be a little more hazy, but it's still not the same thing as a backfiring car.

Sure that's the case when you're firing guns and backfiring cars across an open field from whatever you're measuring sound with but sounds tend to get screwed up when distance, obstructions and white noise are added.

You make it sound like people have never had to do signal analysis before.

Have you seen what NASA can do with images?