General Machine Learning Classifiers and Data Fusion Schemes for Efficient Speaker Recognition

Data fusion methods can take advantage of the concepts of diversity and redundancy to improve system performance. Diversity can be used to improve system performance through the incorporation of different information. Similarly, redundancy can achieve the same goals through the re-use of data. These concepts have been thoroughly applied on pattern recognition problems. The basic idea is that if several classifiers can be constructed, whose errors are mutually uncorrelated, then performance advantages can be obtained through the propel classifiers fusion. The contribution of this paper is to study the fusion of several machine learning classifiers and to analyze data fusion schemes for text independent speaker identification. Feature spaces are defined by combining the Mel-scale Filterbank Cepstrum Coefficients (MFCC) and delta coefficient. Each feature is modelled using the gaussian mixture model (GMM) that constructs a speakers’ models dictionary used later as inputs for classification. Then, four popular supervised machine learning classifiers are considered, namely the multilayer perceptrons classifier (MLP), the support vector machines classifier (SVM), the decision tree (DT) classifier and the radial basis function networks classifier (RBF). The scores (outputs) of classifiers are considered according to different scenario. Results showed that the best performance had been achieved by fusing the SVM, the MLP and the DT classifiers that reported a speaker identification rate equal to 94.15 %.