Underconstrained Stochastic Representations for Top-down Computational Auditory Scene Analysis

Since Bregman published his unifying account of psychological results in auditory organization, Auditory Scene Analysis [1], there has been a series computational models of these principles. The dominant approach, as embodied in the dissertations of Cooke [2], Mellinger [3] and Brown [4], and elsewhere [5], may be characterized as follows: First the sound is processed by a conventional signalprocessing module intended to simulate early auditory processing. This transformation is then converted into discrete elements, and through a search process these elements are organized into different groups corresponding to different ‘sources’ believed to exist in the sound. Within this outline, the models vary considerably in their level of physiological faithfulness, the kinds of sounds they are intended to process, the principles and strategies used for grouping, and the form of final output.