Recognition of phonetic labels of the timit speech corpus by means of an artificial neural network

Abstract

An artificial neural network architecture in terms of Gaussian kernels together with an associated network training algorithm is proposed for phonetic density estimation. Past experience suggests that the success of the network in estimating p.d.f.s depends on how well the densities are represented by the training samples, the number of kernel neurons employed and how thoroughly the network is trained irrespective of the nature of the underlying densities, be it Gaussian or non-Gaussian. This network is now applied to real-life phonetic densities of over one million feature vectors derived from the TIMIT speech corpus. For verification purposes, a classifier based on such density estimations is constructed and its recognition rate is compared with that of a perceptron classifier, a nearest-neighbour classifier and a classifier based on the hidden Markov models exploiting the dynamic information of speech. As a conclusion, the recognition capability of the proposed network is found to be compatible with that of a Bayes classifier.