Combining Harmony-Based and Novelty-Based Approaches for Structural Segmentation

This paper describes a novel way to combine a well-proven method of structural segmentation through novelty detection with a recently introduced method based on harmonic analysis. The former system works by looking for peaks in novelty curves derived from self-similarity matrices. The latter relies on the detection of key changes and on the differences in prior probability of chord transitions according to their position in a structural segment. Both approaches are integrated into a probabilistic system that jointly estimates keys, chords and structural boundaries. The novelty curves are herein used as observations. In addition, chroma profiles are used as features for the harmony analysis. These observations are then subjected to a constrained transition model that is musically motivated. An information theoretic justification of this model is also given. Finally, an evaluation of the resulting system is performed. It is shown that the combined system improves the results of both constituting components in isolation.