I agree with you but I wonder if this algorithm could be reworked to be an important part of a system of that considers context. Roughly, what if we consider x

I agree with you but I wonder if this algorithm could be reworked to be an important part of a system of that considers context. Roughly, what if we consider x_i the sequence of latent chords, and y_i the observed notes. If we reframe the accidental-based sorting this uses to act like a likelihood function e.g. P(y_i | x_i) ~= 1/exp(# accidentals), then we "only" need a transition function P(x_i | x_i-1) which describes what chords typically follow which preceding chords, and then we have an HMM system for which we can do inference easily. You can also imagine extending the order (number of preceding chords regarded as context), or training some higher order latent state z which might carry information about mood or genre and what chord sequences are appropriate in context e.g. P(y) = P(z_0) \Product P(z_i | x_i-1, z_i-1) P(x_i|z_i) P(y_i|x_i)