A probabilistic graphical model for the music harmony similarity task

Methods for measuring music similarity allow for implementations of completely automated content-based music recommendation systems (similar to Pandora, but without the manual work of expert musicologists). This paper presents a novel method of measuring music harmony similarity based on an original probabilistic graphical model. The model includes information about the current chord and mode; we introduce a hidden parameter, style, which governs the probability of using of a certain chord within the context of a certain mode, and propose to measure the similarity as a distance between parameter vectors of the probability distribution function for style. Similar to some methods for extracting chord progressions, our model includes neither the rhythmic information nor the dependencies between neighboring chords. We describe the implementation of our model done with the Infer.NET system and show experimental results on generated data. The results of experiments with real-world data are negative, which indicates that simple bag-of-chords models are not suitable for the music similarity task.