A Combination Approach of Gaussian Mixture Models and Support Vector Machines for Speaker Identification

Gaussian mixture models are commonly used in speaker identification and verification systems. However, owing to their non discriminant nature, Gaussian mixture models still give greater identification errors in the evaluation process. Partitioning speakers database in clusters based on some proximity criteria where only a single cluster Gaussian mixture models is run in every test, have been suggested in literature generally to speed up the identification process for very large databases. In this paper, we propose a hierarchical clustering scheme using the discriminant power of support vector machines. Speakers are divided into small subsets and evaluation is then processed by GMMs. Experimental results show that the proposed method reduced significantly the error in overall speaker identification tests.