The effects are pretty cool, but I can't help but wonder how much better it would be if done with a bit of intention and knowledge of what these algorithms are actually doing.

"We found that while the raw data makes no music, the waveforms share patterns"

Its not that surprising, since music is frequency based while pictures are spatially based. I expect you could get something that sounded like music from a picture if you did some Fourier-transform filtering before with different interval domains (different sizes of 'chunks' being converted to frequencies).