Physiological models of masking and speech processing

Abstract

Auditory models might be applied to speech coding either as objective evaluation measures, or as coder stages in coder/decoder systems. Both applications are based on the premise that these models accurately represent all the information that is relevant to auditory perception. In this talk, two fundamental problems that need to be worked out in order to make auditory models practically useful in speech coding are identified: the choice of a distance measure for comparing model outputs, and the design of the decoder. These ideas are illustrated using a model of peripheral auditory processing that is based on physiological recordings from auditory-nerve fibers, and predicts psychophysical performance in masking and discrimination for a variety of stimulus conditions. In this model, intensity difference limens in restricted regions of the frequency spectrum depend on overall spectral shape and sound level, so that the coding accuracy in different frequency bands would have to vary with time. Model outputs will be compared for natural and coded speech, and the role of nonlinear phenomena in the inner ear (such as two-tone suppression ) will be discussed. [Work supported by NIH Grant NS13126.]