Okay, so I’m a software engineer and one of my pastimes is keeping an ear open to various Boulder County public safety dispatch channels (that’s what saved my butt back in December 2021). I’ve been working on a system for myself to do computer transcriptions of these dispatch channels just so I can get alerted when they say certain keywords., e.g., if they mention my kids’ schools. But relying on computers to interpret radio traffic is iffy at best.

First of all the voice transcriptions are horribly wrong most of the time. Here’s an example of a BVSD security transcription: “can you show me a metal arc for their ligandro, please?” Now BVSD security, as do almost all other public safety agencies, have a pretty limited radio vocabulary for the sake of clarity, but the voice transcription models aren’t trained on that, so if you haven’t actually listened to hours of their chatter, you would have no idea that they were actually saying, “can you show me at Meadowlark for their lockdown drill, please?” (Also… WTF is a “ligandro?!”)

Next, AI, at least for the moment, has very little if not no capability to question its own assumptions with any form of logic or reasoning in order to ask itself if that even makes sense. Believe me, I wish it did because then my homemade project would be a lot better.

So for now I just have to hope that it gets a word or two right enough that I can get a notification and tune in myself to figure out what is going on. This is probably why Watch Duty is a little slow to respond but very accurate: they are staffed by real humans who have listened to countless hours of radio traffic.

The human brain is very adaptable. The computer “brain” is not.

P.S. Anyone who is interested can watch these near-real-time so-called “transcriptions” here.

This post is partially in reaction to this Daily Camera article.