COMPARISON OF VARIOUS ADAPTATION MECHANISMS IN AN AUDITORY MODEL FOR THE PURPOSE OF SPEECH PROCESSING

This paper presents preliminary results from the comparison of a number of existing models for the inner hair cell/synapse region of the auditory periphery, for the task of speech processing. Particular emphasis is placed on the models' ability to enhance dynamic characteristics of the speech signal, especially abrupt changes in amplitude, as reflected in the adaptation properties of each model. Such characteristics could be useful for the task of phonetic segmentation. Three adaptation models were chosen, and they were each combined with a computational model of basilar membrane mechanics for the purpose of generating responses. The responses of the basilar membrane/adaptation model combinations are compared with the responses which were obtained with a simple envelope-detecting type of inner hair cell model. In addition, an alternative model for adaptation is proposed. This model is computationally very simple, and is optimised for the task of enhancing dynamic aspects of the speech signal, rather than attempting to faithfully reproduce physiological data, as is the case with most other models. It is intended that a model of the type proposed in this paper could be useful for real-time implementation in a front-end processor for a speech recognition system.