EUROSPEECH 2003 - INTERSPEECH 2003
Despite all advances in the speaker recognition domain, Gaussian Mixture Models (GMM) remain the state-of-the-art modeling technique in speaker recognition systems. The key idea is to approximate the probability density function ( pdf) of the feature vectors associated to a speaker with a weighted sum of Gaussian densities. Although the extremely efficient Expectation-Maximization (EM) algorithm can be used for estimating the parameters associated with this Gaussian mixture, there is no explicit method for predicting the best number of Gaussian components in the mixture (also called order of the model). This paper presents an attempt for determining the "optimal" number of components for a given feature database.
Bibliographic reference. Arcienega, Mijail / Drygajlo, Andrzej (2003): "On the number of Gaussian components in a mixture: an application to speaker verification tasks", In EUROSPEECH-2003, 2673-2676.