A computational approach to the dynamic aspects of primitive auditory scene analysis

Abstract

Recent psychophysical and physiological studies demonstrated that auditory scene analysis (ASA) is inherently a dynamic process, suggesting that the system conducting ASA constantly changes itself, incorporating the dynamics of sound sources in the acoustic scene, to realize efficient and robust information processing. Here, we propose computational models of ASA based on two computational principles of ASA, namely, separation in a feature space and temporal regularity. We explicitly introduced learning processes, so that the system could autonomously develop its selectivity to features or bases for analyses according to the observed acoustic data. Simulation results demonstrated that the models were able to predict some essential features of behavioral properties of ASA, such as the buildup of streaming, multistable perception, and the segregation of repeated patterns embedded in distracting sounds. © Springer Science+Business Media New York 2013.