NN-based automatic sound classifier for digital hearing aids

This paper centers on exploring proper training algorithms for multilayer perceptrons (MLPs) to be used within digital hearing aids. One argument usually considered against the feasibility of neural networks on hearing aids consists in both their computational complexity and the hardware constraints the hearing aids suffer from. Within this framework, this work focuses on studying the influence of a number of training methods for an MLP able to automatic classify the sounds entering the hearing aid into three classes: speech, noise and music. The training methods explored are Gradient Descent, Levenberg-Marquardt, and Levenberg-Marquardt with Bayesian Regularization. Our results show how the proper selection of the training algorithm leads to a good mean probability of correct classification of 91.7% along with a low number of neurons, the computational complexity being thus reduced. These results have been successfully compared to those obtained from a k-Nearest Neighbors algorithm, which exhibits poorer performance.