regarding the problem of ornamentation, variation, and notation:

If I were to attempt a computational classification I'd probably start with an impoverished representation that omits ornaments but preserves variability within singer (from verse to verse and from singing to singing) as well as across singers. So I'd have to rely on an expert musician to decide which notes are ornaments, and extract the pitches and durations that should be considered the underlying 'melody'. A theory of ornamentation could be layered on later. Other than note bending (which could be considered an ornament), mathematically it's no problem to define both pitch and duration of each note on an analog scale without reference to any theory of tuning or time signature. But again expert judgement might be valuable, for example in discerning whether variations in the pitch of a note in the melody in different verses or repetitions or persons is attributable to pitch inaccuracy ("noise") or drift in the singer's implicit key ("nonstationarity"), versus a meaningful modulation ("signal"). I suppose one could postulate that any drift is accidental, and impose a global correction so that the tonic note remains stable for the duration of the song; and then compare the pitch of the same note in the tune from singing to singing to determine what the "intended" pitch is for that singer, and whether any deviations from verse to verse are systematic or random. My guess is that variations in the note duration are usually either text-driven or expressive and therefore more likely to be consistent from singing to singing.... but this is turning into more of a PhD thesis, and I don't have a PhD student just a bunch of Freshman undergraduates.

I suppose the advent of digital recording, pitch analysis software, music editing software and computer science/artificial intelligence make possible today an analysis that would've been impossible for S.P. Bayard et. al. So maybe it hasn't been done.   Interesting.

Pamela