Musical Feature and Novelty Curve Characterizations as Predictors of Segmentation Accuracy

Abstract

Novelty detection is a well-established method for analyzing the structure of music based on acoustic descriptors. Work on novelty-based segmentation prediction has mainly concentrated on enhancement of features and similarity matrices, novelty kernel computation and peak detection. Less attention, however, has been paid to characteristics of musical features and novelty curves, and their contribution to segmentation accuracy. This is particularly important as it can help unearth acoustic cues prompting perceptual segmentation and find new determinants of segmentation model performance. This study focused on spectral, rhythmic and harmonic prediction of perceptual segmentation density, which was obtained for six musical examples from 18 musician listeners via an annotation task. The proposed approach involved comparisons between perceptual segment density and novelty curves; in particular, we investigated possible predictors of segmentation accuracy based on musical features and novelty curves. For pitch and rhythm, we found positive correlates between segmentation accuracy and both local variability of musical features and mean distance between subsequent local maxima of novelty curves. According to the results, segmentation accuracy increases for stimuli with milder local changes and fewer novelty peaks. Implications regarding prediction of listeners’ segmentation are discussed in the light of theoretical postulates of perceptual organization.