> There wasn't any limitation from the mechanical perspective. The AI did a nearly perfect job of recognizing notes. I don't know exactly why it missed some notes, but my hypothesis is that it was the speed of notes in quick succession. If I do a follow up video with Jon Bot Jovi, I'll dig into it more!
a couple of nuances make the remaining 3% a bit more difficult to get, two that I noticed is that the bot is bad at holding sustain notes and also tends to make mistakes during long hammer-on sections because it hits two notes at the same time instead of realizing that they are individual notes.
the second one is probably easy to fix by spreading the notes out a bit (hyperspeed mode) but the first one is a bit more annoying to program, probably
Agreed the sustained notes would be a bit of a programming annoyance, but if I had spent more time, I think I could have cracked it. I did get some MVP working for it but it was throwing some other things off, so I abandoned it in an effort to maximize accuracy.
I think it might be easy implementing sustain on any and all notes until the next note needs to be hit. So instead of press-release-wait you do press-wait-release.
One thing was a bit unclear: is Raspberry Pi doing the blob detection, or is it running on the Macbook?
Also, why only 97%?