I guess it probably didn't go off much tonight.

Go Habs go :)

Really cool.

What are you doing for away games? Can it still detect Habs goals? You mentioned you had to train the algorithm not to detect the opponents' goal horn. Couldn't you just scrape the schedule to detect if you're playing at home or away and use a different model for each?

It's really about the comentator's yell, not the goal horn. So as far as he's excited for Habs goals and not too excited for the other team's goals, it works fine.

You watch the French commentary, I presume.

Yup, the current model is trained on the TVA sports broadcast

Yes, the English broadcasters at least pretend to be neutral.

Go Sens Go!

Pretty cool my man but how do you deal with false positives?

I don't know about hockey but in soccer (football), commentators are as much excited about near-goal scenarios like the ball hitting one of the posts or the goalie making a superb save as goals themselves.

I can't imagine employing such a system for football games. What do you think?

For false positives during a live broadcast, I have the big USB button ready to cancel the light show and goal song. I did not mention in the blog post is that it stops the goal song by playing the Fail Trombone sound (https://www.youtube.com/watch?v=iMpXAknykeg) when you use the USB button to cancel light show

Really not sure how it would work for other sports. The only reason it worked so well is probably because they GOOOAAAAAALLLLLLLL yell is long and distinctive versus everything else. It's not about intensity as much as a pattern in the frequency spectrum. I think it's worth a try with soccer and football.

Did you consider trying to use speech recognition rather than spectrogram analysis?