First, some general info.

Here’s a map of the area that will be monitored by Shotspotter. The town has signed up for a year long trial, with a total cost to the town being just under $200k.

Shotspotter is giving the town a 25% discount for the trial (3 months free!) so if the town were to re-up the annual cost would increase (and Shotspotter would surely encourage the town to expand the coverage area beyond this small section of east durham, their original proposal covered a much larger area).

The city maintains a portal with info about the program here: https://www.durhamnc.gov/4827/ShotSpotter-Durham

There's a lot of good coverage already about how Shotspotter is very bad at what it does, is harmful to communities where it is deployed, and has worked with police in the past to fabricate gunshot evidence.

Greg’s got all that covered in this epic thread: https://twitter.com/greg_doucette/status/1511826946790993927?s=20&t=W65q1v5GiBmuIZ-uKDB17Q

There aren’t a ton of companies doing noise recognition. It relies on a lot of the same underlying tech as computer vision, but is more complex due to the way sound changes as it travels through air and bounces off surfaces.

There are fewer commercialization paths compared to CV… and security tends to be the starting point for most companies working in this domain.

Detecting and localizing gunshots isn’t easy, but it is easier than many other sound detection problems because they're so loud.

Shotspotter was founded in 1996, before the deep-learning revolution that brought high accuracy rates to sound detection.

There isn’t much info about their tech, but considering that every single detection event must be reviewed by a human prior to being passed on to their customers, I’m guessing it is a little dated.

I’m sure they’d argue that this is to ensure that police are only sent valid detection events, but to me it reads like they don’t trust their tech enough to fully automate it.

This human review layer is critical to how they structure their SLA or service-level agreement.

This defines the minimum level of service they’re contractually obligated to meet, and I think sheds a lot of light on why Shotspotter is so problematic.

The first says they’ll detect 90% of actual gunshot events. The second says that 90% of detections will be reviewed by a human within a minute of their detection.

The third is standard for any SaaS product and promises three 9’s of uptime, or that their system is allowed to be unavailable (due to unplanned technical issues) for up to 9 hours a year.

Notably absent here is any kind of promise about limiting false positives.

They are obligated to detect 90% of real gunshots, but they aren’t penalized if they mislabel a firecracker or backfiring car as a gunshot.

Further, they have to label quickly, with little context (they are only allowed to listen to 1 sec before/after the triggering sound).

This SLA is engineered to encourage over-reporting. When in doubt, the human reviewer will "play it safe” and label as gunshot.

This means more cops showing up to benign events with a “shots-fired” pretext. Not good.

There are other companies that make sound detection their business.

Audio Analytic (based in the UK) won’t ship a detection algorithm unless they’ve achieved at least 98% accuracy and that measurement includes false positives.

Speaking of accountability, there isn’t much here.

What happens when Shotspotter doesn’t meet these low standards? Well Durham gets a small discount on their next year of service (if they choose to renew)!

But that’s only if Shotspotter significantly misses on 2 of the 3 SLAs, so as long as they keep their support staff on their toes and their website online the 90% detection thing doesn’t even really matter.