Effective features to define harmonic distance models

Tonal Pitch Space (TPS) has been proposed as a harmonic distance model, which provides a way to compute numerical distances between two chords with keys. Although the theory itself seems sound, it still contains arbitrary presuppositions in its definition, and its adequacy has not well been investigated. In this study, we propose a framework in which we construct, train, and evaluate a variety of distance functions that share the same domain and range as TPS. We define three basic harmonic features (mode, tonic, and degree), and utilise these to define various distance functions by combining them, and then we evaluate all simple combinations exhaustively. We then train and evaluate each function through the task of key estimation. This entire process is to provide a perspective from which to examine how well various features and their combinations can represent harmony through our distance models. Experiments confirmed that some aspects of the TPS are indeed well designed, but also showed that its performance can be surpassed with a considerably smaller number of parameters

Acoustics · Harmonic · Physics · Acoustic Wave Phenomena Research · Computer Science · Music and Audio Processing · Speech and Audio Processing · Artificial Intelligence