Approach to speaker identification using DP-matched LVQ neural networks

Abstract

Automatic speaker identification is the other aspect in speech science, vis-a-vis spoken word recognition. In its traditional realization as a pattern classification task, there are some difficulties, such as the uncertainty of personalities, the dynamic variation of features, and so on. Nevertheless, vector quantization based pattern classifier still gains remarkable interest, because of its simplicity and the recent development in parallel computation technology. There are several reports on the combination technique of LVQ and other method for improvement of the conventional classifier. However, this paper concentrates on a hybrid algorithm of LVQ and DP-matching. The feature normalization capability in both the time and frequency domains of such a method can decrease the incorrect speaker identification which is caused by the variation of feature vectors in a short-term or a long-term. The experiment of 10 speakers identification shows that the proposed method can produce the better reference vectors, and hence it can promote the correct identification.