Detection of telephone-quality speech using radial basis function networks

This study seeks to address the problem of detecting the presence of human speech in a short-time segment of a telephone line signal. It consists of the design, implementation, and evaluation of a pattern classifier based on a neural network paradigm to identify speech over a background of silence and other non-speech signals. The algorithm for speech detection is based on the method proposed by J. Hoyt and D. Wechsler (1994) with modifications to process a more suitable set of signal feature measurements. The classifier uses radial basis function (RBF) networks to make this a two-class decision problem. The study experimented on two different basis functions for the hidden layer nodes of the RBF network to determine the effects of using the Mahalanobis and Euclidean distance on the accuracy of detection of the designed classifier