jagiii

I'm still not perfectly happy with the audio response, tbh. But it mostly suffices.

I experimented with both approaches, and landed on mixing the band energy normalized by global energy vs. per-bin max energy, with a very arbitrary coefficient that seems to work. I also keep a few copies of the levels per band with different rates of decay, so I can add features in the sketches which should be sensitive to transients, vs. features which can rely on slower rates of change.

Not at all an expert in any of this and it was a lot of trial and error, still lots of room for improvement.

I had this concern at first, but using a glue stick seemed fine. The gluing doesn't have to be extensive either, since the paper gets clamped between the plastic parts when I screw it together.

I took the very thin 3D printed cell outline shape, a grey screenshot in my post, and used it to trace the mulberry paper. Then cut and glued paper onto the bottom of that cell outline, and screwed the outline plate into the main assembly.

This is by far one of the biggest areas where I could improve the design. I'm not happy with how fragile it is. The good thing about using just paper screwed from the top is it's thin enough to prevent light bleeding between cells.

author here, good digging! I added the MIT license.

Getting 5-fold symmetry to work was tricky, but the repo you found has the processing sketch where I figured it out. I basically took the point generation process I described and rotated it upon itself 5 times. I tested this with different radial symmetry numbers actually, and 5 also happened to look the best and also was convenient for manufacturing the PCBs.

Then I had AI add the ability to rotate a bounding rectangle for one 'quintant' around the shape, and found the division of cells which minimized the area, since PCB cost scales with that.