FRIHA, Souad University of Tebessa - Institut de Génie Electrique, ALGERIA , MANSOURI, Nora University of Mentouri - Laboratoire d Automatique et de Robotique, ALGERIA , TALEB AHMED, Abdelmalik Université de Valenciennes et du Hainaut Cambrésis - LAMIH UMR CNRS-UVHC 8530, France

THE INFINITE GAUSSIAN MODELS: AN APPLICATION TO SPEAKER IDENTIFICATION

21553

When modeling speech with traditional Gaussian Mixture Models (GMM) a major problem is that one need to fix a priori the number of GMMs. Using the infinite version of GMMs allows to overcome this problem. This is based on considering a Dirichlet process with a Bayesian inference via Gibbs sampling rather than the traditional EM inference. The paper investigates the usefulness of the infinite Gaussian modeling using the state of the art SVM classifiers. We consider the particular case of the speaker identification under limited data condition that is very short speech sequences. Basically, recognition rates of 100% are achieved after only 5 iterations using training and test samples less than 1 second. Experiments are carried out over NIST SRE 2000 corpus.

101

Speaker Identification , Infinite GMM , SVM , Dirichlet Process , Gibbs Sampling

Courrier Du Savoir

107