Effective Features for Finding Chord Interpretation Paths

Abstract

We humans naturally feel tonality when we hear music, but its actual mechanism is still unknown. There have been many approaches to understand this mechanism, but one promising method is to measure the distance between chords, as was proposed in Tonal Pitch Space (TPS). Although the theory seems theoretically convincing, we could not know which features contribute most and which others are useless in an objective manner. In this study, we try to define a new distance model by clarifying the effective set of features among them. Based on the principle of the shortest distance, which is claimed by TPS, we estimate each distance by optimizing features to obtain the most plausible chord interpretation (i.e., a degree/key pair for a chord name). We propose the set of functions based on these features and we evaluate all simple combinations of them exhaustively. It turns out that accuracies almost max out with around 20 effective parameters in the functions, achieving about 80-90% accuracy, that outperforms the original TPS.