Comparing SVM and GMM classifiers on the parametric feature-sets

State of the art speaker identification systems use the Gaussian mixture models (GMM) classifier. Support vector machines (SVM) offers a competing classification algorithm. Both classification methods have been evaluated on speaker recognition tasks and have shown to produce uncorrelated errors with sometimes similar performance. In this paper their performance is compared on the amount of data required for optimal performance and effect of different spectral nonlinearities in the feature-sets. The results show that for limited training data the SVM classifier is better but as the data is increased the GMM classifier outperforms it. The SVM classifier is also very expensive. This suggest that the GMM classifier will continue to be the popular classification engine for speaker recognition tasks. SVM can be used in special cases such as when there is limited data or as a secondary classifier, otherwise it is computationally very expensive and offers little in return.