The performance of a novel pre-processor for the learning vector quantisation classifier

Abstract

In this paper a novel pre-processing technique, known as the pseudo-cepstrum, is described. The technique is applied to a speaker independent, continuous speech recogniser, using Learning Vector Quantisation (LVQ) for the classification stage. The results of a set of performance comparisons between the pseudo-cepstrum, the conventional log-cepstrum and linear prediction will be presented. For each of these three pre-processors the effect of using a mixture of instantaneous, and first and second time differences are investigated. It will be shown that the pseudo-cepstrum, using a mixture of instantaneous and first time differences gives superior results. When this is translated into the lip shape classification, which is the intended output of this recognition system, the results are further improved.