An ADPCM architecture based on models of the auditory system

A class of ADPCM speech predictor architectures that are based on models of the human auditory system is developed. The transfer function of the predictor is of the pole/zero type with fixed poles and adaptive zeros. Choice of the fixed poles is made using auditory models, in contrast with the standard trial-and-error techniques. A state-space formulation of the prediction filter provides the connection between the standard fixed pole/adaptive zero models and the transversal filter model of the auditory system. Two realizations of the prediction filter are described: a parallel realization, which directly uses the auditory models, and a cascade realization, which allows direct manipulation of the poles and zeros of the filter. The gradient search method is applied to identify the filter coefficients