Application of combining classifiers for text-independent speaker identification

Speaker recognition systems usually need a feature extraction stage witch aims at obtaining the best signal representation. States of the art, speaker identification systems are based on a cepstral feature extraction follow by an individual classifier or a hybrid classifier. Nowadays, an alternative approach consists in fusing different features or different classifiers are increasingly used. In this paper, different features are defined. Each feature is modeled using the Gaussian mixture model and construct a speakers´ models dictionary. These dictionaries are used by the multilayer perceptron (MLP) classifier, the support vector machines (SVM) classifier and the decision trees (DT) classifier for matching and the scores (outputs) of all classifiers are then considered for combination. Results indicate that the use of combining classifiers with different features is an effective way to attack the problem of text-independent speaker identification.