Probabilistic graphical models of harmony music analysis: a survey

This paper presents the current state of art in the field of automated musical harmony analysis. Research in this field can be motivated by the real-world problems of creating completely automated content-based music recommendation systems (similar to Pandora, but without the manual work of expert musicologists). The paper is mainly focused on probabilistic graphical models as one of the most promising approaches, although we also give background in alternative methods. We consider works that use Markov chain models, hidden Markov models, and multi-level graphical models. Along with the models that capture only harmonic information – chord progressions, in some cases also the key, – we also list several models that combine harmonic structure with rhythmic or stream structure.