Learning Tonal Harmony from Bach Chorales

Tonal harmony analysis is an intriguing cognitive skill, combining general domain knowledge with contextual cues. In this work we cast it to a Supervised Sequential Learning problem (SSL), and introduce the system geraint, showing how such a problem can be solved via the HMPerceptron algorithm. We explain the main concepts used in music analysis, their use within an automatic environment, and provide their cognitive motivation, such that geraint is both effective on a computational viewpoint, and justifiable on musical and psychological accounts. geraint’s predictions are evaluated on a corpus of 4-parts harmonized chorals by J.S. Bach. We report on the experiment, and discuss our system’s results in comparison to literature.