This doesn't play MP3s.

It just plays notes; those notes are sourced from MIDI files (though they could also be sourced from a MIDI keyboard or similar, so a human can play the machine live).

The MIDI data stream, at it's most simple basis, is just a series of "note on" and "note off" commands. (MIDI can additionally do a lot more than this, but let's not dwell on that.)

This concept actually compares particularly well with what a punched paper roll accomplishes: That, too, is just a series of "note on" and "note off" commands. For every possible note or drum or percussive (or automaton) thing the machine playing the paper roll can do, there's either a hole in the paper ("on") or there is not ("off").

One system is digital and happens in numberland with circuits and/or code, while the other is pneumatic and uses valves and pipes and pumps to get the work done.

But they're both binary systems, so it can be pretty straight-forward to convert between the two.

Relatedly: A somewhat aloof chap in England has found himself with a fondness for pipe organs. He scored a whole church organ from a lady's house, converted it to MIDI with a rather grand assortment of custom PCBs and rewiring, and put it in his museum. (That organ was designed to use electricity and solenoid valves, and meant to be played live instead of with a paper roll, but it's the same game: Binary is binary.)

The process is documented here: https://www.lookmumnocomputer.com/joans-church-organ