AUDITORY DISTORTION MEASURE FOR SPEECH CODER EVALUATION - HIDDEN MARKOVIAN APPROACH

This article introduces a methodology for quantifying the distortion introduced by a low or medium bit-rate speech coder. Since the perceptual acuity of a human being determines the precision with which speech data must be processed, the speech signal is transformed onto a perceptual-domain (PD). This is done using Lyon's cochlear (auditory) model whose output provides the probability-of-firing information in the neural channels at different clock times. In our present approach, we use a hidden Markov model to describe the basic firing/non-firing process operative in the auditory pathway. We consider a two-state fully-connected model of order one for each neural channel; the two states of the model correspond to the firing and non-firing events. Assuming that the models are stationary over a fixed duration, the model parameters are determined from the PD observations corresponding to the original signal. Then, the PD representations of the coded speech are passed through the respective models and the corresponding likelihood probabilities are calculated. These probability scores are used to define a cochlear hidden Markovian (CHM) distortion measure. This methodology considers the temporal ordering in the neural firing patterns. The CHM measure which utilizes the contextual information present in the firing pattern shows robustness against coder delays.