Effects of musicianship and experimental task on perceptual segmentation

Abstract

The perceptual structure of music is a fundamental issue in music psychology that can be systematically addressed via computational models. This study estimated the contribution of spectral, rhythmic and tonal descriptors for prediction of perceptual segmentation across stimuli. In a real-time task, 18 musicians and 18 non-musicians indicated perceived instants of significant change for six ongoing musical stimuli. In a second task, 18 musicians parsed the same stimuli using audio editing software to provide non-real-time segmentation annotations. We built computational models based on a non-linear fuzzy integration of basic and interaction descriptors of local musical novelty. We found that musicianship of listeners and segmentation task had an effect on model prediction rate, dimensionality and components. Changes in tonality and rhythm, as well as simultaneous change of these aspects were important to predict segmentation by listeners. Our results suggest that musicians pay attention to more features than non-musicians, including more high-level structure interactions. Prediction of non-real-time annotations involved more features, particularly interactions thereof, suggesting high context dependency. The role of interactions on perception of musical change has an impact on the study of neural, kinetic and speech stream processing.