Adaptive environmental classification system for hearing aids

Abstract

This thesis develops an adaptive environmental classification system for hearing aids. Two types of classifiers, minimum distance and Bayesian, are modified to include an adaptive layer that allows the classes to split or merge based on changes in the environment. In order to test the adaptability to the environmental change, both systems were first trained using two classes: Speech and Noise, followed by a testing period where, in addition to Speech and Noise, samples from a third class Music were introduced. Both systems were successful in detecting the presence of the new class Music and estimated its parameters obtaining an accuracy that is as high as the accuracy obtained through a non-adaptive supervised learning. In addition, the accuracy for the merging algorithm in both systems also met that of a non-adaptive system. In comparing the two systems, the adaptive Bayesian classification system resulted in a higher accuracy in classifying the environment into the three classes: Speech, Noise, and Music, following the adaptation process.