Multi-scale Modelling of Segmentation

While listening to music, people often unwittingly break down musical pieces into constituent chunks such as verses and choruses. Music segmentation studies have suggested that some consensus regarding boundary perception exists, despite individual differences. However, neither the effects of experimental task (i.e., real-time vs. annotated segmentation), nor of musicianship on boundary perception are clear. Our study assesses musicianship effects and differences between segmentation tasks. We conducted a real-time experiment to collect segmentations by musicians and nonmusicians from nine musical pieces. In a second experiment on non-real-time segmentation, musicians indicated boundaries and their strength for six examples. Kernel density estimation was used to develop multi-scale segmentation models. Contrary to previous research, no relationship was found between boundary strength and boundary indication density, although this might be contingent on stimuli and other factors. In line with other studies, no musicianship effects were found: our results showed high agreement between groups and similar inter-subject correlations. Also consistent with previous work, time scales between one and two seconds were optimal for combining boundary indications. In addition, we found effects of task on number of indications, and a time lag between tasks dependent on beat length. Also, the optimal time scale for combining responses increased when the pulse clarity or event density decreased. Implications for future segmentation studies are raised concerning the selection of time scales for modelling boundary density, and time alignment between models.