A hybrid GMM-SVM speaker identification system

This paper proposes a system that combines the power of generative Gaussian mixture models (GMM) and discriminative support vector machines (SVM) in a speaker identification task. The classification methods are different and they also exhibits uncorrelated errors and this is used to improve performance of the speaker identification system. Whereas GMM needs more data to perform adequately and is computationally inexpensive, SVM on the other hand can do well with less data and is computationally expensive. A system where SVM post processes the results of a GMM system is proposed and it is shown that it is able to reduce speaker identification errors by over 11% on a database with 630 speakers. Similar hybrid systems have been proposed before but this is unique since both classifiers use the same feature vectors. Improved performance is found by using optimal parameters (σ, C) for the SVM Gaussian kernel.