Hacker News Reader

Comment by jaywalk | Hacker News Reader

You say falsification of evidence, they say manual review by an "expert" to correct errors from the automated process.

Reclassifying the type of sound is possibly reasonable. Changing the coordinates seems like falsification of evidence. Later in the article it mentions adding a fifth shot in a different event.

If shotspotter wants to be relied on for evidence, allowing technicians to go back and modify data seems like a recipe to quickly eliminate its use in cities and destroy its market share. Adding notes about manual review after the fact would be the only acceptable thing, but the originals should never be modified/deleted.

Sorry, but I’m an expert shooter. I can reasonably identify the round and sometimes the gun on the range next to me - I would say I have an excellently tuned ear for these things…

And not a chance I could do that based on microphone playback between firework and gunshot with some random mic in a city. 1/2 the “trick” is knowing the location, at my range I know the echo of a 12ga vs a 45ACP, the crack difference in a 223 or a 6.5prc, if I went to some random place and you shot one random shot, it would be ridiculously less accurate.

Shotspotter seems like a ton of bullshit to me after reading this article including…

> Both the company and the Rochester Police Department “lost, deleted and/or destroyed the spool and/or other information containing sounds pertaining to the officer-involved shooting,”

> And not a chance I could do that based on microphone playback between firework and gunshot.

I’m happy to agree with you on that point, since I don’t have the experience there.

My larger point is that I could reasonably say allow for shotspotter to review the sound, but the location I’m assuming is pretty darn accurate.

And yes, the article completely destroys their credibility as a reliable source of evidence.

Multipath effects, which for sound are huge in an environment with large flat surfaces, could create phantom origin spots at the fringes of the covered area.

So if your point is it’s too complex and variable for automatic processing; how does adding a human arbitrarily changing classification and location going to improve accuracy? Unless it’s the shotspotter tech put there shooting people, seems the company is very successful bullshit either technically or ethically.

If there was solid external evidence of when a shooting occurred (eg gps timestamped video of the event) I could see them doing it purely for calibration purposes.

Clearly those adjustments should not be admitted as evidence in a court case. ShotSpotter just needs to take the 'L' and they will have to prosecute without (or admit whatever was originally recorded as corroborating evidence).

>but the location I’m assuming is pretty darn accurate.

I remember reading about a similar technology used by soldiers in Afghanistan to find snipers, but Afghanistan is largely an open area. I don't see how it would be accurate when the acoustics of multiple buildings are in effect.

The idea is attractive because it's quite obvious how well ShotSpotter would work in a cleanroom environment. The part that's less intuitive is how far the standard deployment environment is from that standard and what margin the product has to accommodate.

If the source audio is not reviewable in court, none of it should be admissible as evidence. Algorithms are just laundering bad opinions like blood splatter analysis, so that there’s not even a personal reputation at stake.

"Liability laundering" does roll off the tongue more easily than "machine learning algorithm".

>Algorithms are just laundering bad opinions like blood splatter analysis

You have way too much faith in the legal system. So many pseudo-scientific classes of evidence exist:

- Field sobriety tests

- Ballistic forensics

- Blood spatter analysis, as you mentioned

- Bite pattern analysis

- Burn / arson forensics

- Polygraph

- Biometrics under less-than-perfect conditions (your fingerprint reader on your phone works well because it's near perfect conditions. Crime scene partial finger prints are usually insufficient. Same for facial recognition -- works with your iPhone FaceID sensor, doesn't work with the gas station 480x320 camera)

- Most DNA and hair analysis

Pretty much all hocus pocus. Courts don't care. Anything for the conviction. Anyone hoping for fair scientific analysis in a court of law today is in for a rude awakening.

To be fair, polygraph tests aren't admissible as evidence in court, but I completely agree with your larger point. You also left out drug sniffing dogs, which are just a prop for the justification of an otherwise illegal search, and apparently drug sniffing cops too. There was recently a case where a cop claimed in court to have smelled unburned pot from several cars away while driving on the interstate. This was fortunately thrown out but it's not guaranteed.

I agree with you completely, except for the idea that I have too much faith in the legal system.

> manual review by an "expert"

Sounds like that should be required before the results of automated processes can be used as evidence.

I'm sure it is, but the point here is that the manual review appears to be "aided" by the police "suggesting" where they believe the sound came from. And then, wow, wouldn't you know it! The manual review agrees. Case closed.

Reply on news.ycombinator.com