Some aspects of modelling segmentation processes

Introduction The world we live in is continually changing. The images formed on the retina in vision are a sequence of varying, shifting intensities. In hearing, likewise changes in frequency and intensity are recognised to form the basis of our detecting the underlying structure of events in the world. Research in music perception has identified the significance of change in the segmentation of auditory information (Tenney & Polansky, 1980; Deliège, 1987). However, models of how these changes are detected, encoded and processed in melodic listening, are relatively incomplete. Since segmentation relies on our ability to detect and resolve details of complex sounds, it is doubtlessly constrained by both the peripheral and central auditory processing mechanisms (Watson et al, 1990). Also, under certain conditions, our ability to detect changes involves learning (Watson et al, 1990; Leek & Watson, 1984). For example, experiments on frequency discrimination by musicians and non-musicians (Spiegel and Watson, 1984), indicate training as one possible source of performance improvement. Experiments on grouping rules in listening to music (Deliège, 1987) indicate the influence of context (i.e. the effects of stimulus uncertainty) on the performance of non-musicians and musicians. Again, this implies that segmentation is to some extent learned. Between the peripheral processing of a melodic input and 'hearing' its structure, there is a relatively unknown area of functional signal propagation through the auditory echoic store. Advancing our understanding of these processes should provide valuable insights into how an underlying segmentation mechanism may affect melodic listening. We have proposed one possible mechanism of auditory signal propagation applicable to melodic listening which we will briefly outline here (Serman, 2001). The emphasis of this paper is to illustrate the model's behaviour through three examples. Model Since auditory perception involves the identification of changes over a range of temporal scales, the mechanism that we have adopted for modelling signal propagation exploits the idea of multi-scale filtering developed in theories of visual perception and machine vision. The mechanism for multi scale signal decomposition (Marr, 1982) was first adapted for modelling auditory processing by Todd (Todd, 1994). We have used a different way of temporal signal decomposition, a model based on leaky integrator cascades (Serman, 2001). Based on the idea of separate processing of sensory information in auditory cortex (Semal and Demany, 1991), each of melodic descriptors-pitch, loudness and timbre, represents an input to a separate cascade. MusicTracker software estimates these descriptors from the results …