Data-driven, memory-based computational models of human segmentation of musical melody

When listening to a piece of music, listeners often identify distinct sections or segments within the piece. Music segmentation is recognised as an important process in the abstraction of musical contents and researchers have attempted to explain how listeners perceive and identify the boundaries of these segments. The present study seeks the development of a system that is capable of performing melodic segmentation in an unsupervised way, by learning from non-annotated musical data. Probabilistic learning methods have been widely used to acquire regularities in large sets of data, with many successful applications in language and speech processing. Some of these applications have found their counterparts in music research and h ave been used for music prediction and generation, music retrieval or music analysis, but seldom to model perceptual and cognitive aspects of music listening. We present some preliminary experiments on melodic segmentation, which highlight the importance of memory and the role of learning in music listening. These experiments have motivated the development of a computational model for melodic segmentation based on a probabilistic learning paradigm. The model uses a Mixed-memory Mar kov Model to estimate sequence probabilities from pitch and time-based parametric descriptions of melodic data. We follow the assumption that listeners' perception of feature salience in melodies is strongly related to expectation. Moreover, we conjecture that outstanding entropy variations of certain melodic features coincide with segmentation boundaries as indicated by listeners. Model segmentation predictions are compared with results of a listening study on melodic segmentation carried out with real listeners. Overall results show that changes in prediction entropy along the pieces exhibit significant correspondence with the listeners' segmentation boundaries. Although the model relies only on information theoretic principles to make predictions on the location of segmentation boundaries, it was found that most predicted segments can be matched with boundaries of groupings usually attributed to Gestalt rules.