> I suspect naively it shouldn't be too hard to decode an existing recording of music into the notes played.

What you expect not to be too hard to do, is actually the 'holy grail' of the field - Algorithmic Melody Transcription is the kind of thing people write entire papers on[0].

[0]: https://www.ee.columbia.edu/~dpwe/pubs/PolEFGSO07-meleval.pd...

There are a few toy examples if you search "algorithmic music transcription github," one of the better-known ones is [1] (but it's limited to piano music). They're all pretty nascent. It's a real hard problem :)

[1]: https://github.com/rachhshruti/automatic-music-transcription

I recently stumbled across some videos where a producer breaks down how some very good songs are produced. Here's one he does for a Max Martin track where you get screenshots of the track in Pro Tools:

https://www.youtube.com/watch?v=cBDJDZAo41c

> From there, it should be possible to mangle it into a tabulature?

A produced pop track might be a little extreme, but there's a lot going on in these tracks. Even if you know the chord progression, it's hard to tell if a note played by a guitar is played by the rhythm guitar, lead guitar, an overdub, something with double tracking, or even an overtone from the bass, especially when the note makes the chord an add4. The same goes for walking basslines played by the guitar.

You also arrange things differently for solo performances, live performances with a band, and produced performances, so even if you could extract a rhythm guitar tab from a recording, that might not be how you'd want to play it solo.

Celemony Melodyne has supported some form of Audio to MIDI translation for over a decade and may be the most accurate at it.

Back then:

- https://youtu.be/8zeLKMHxbN8?t=561

- https://www.youtube.com/watch?v=OJ2lWslJVOc

Now:

- https://www.youtube.com/watch?v=cvBYZPlitWE