Evaluation of GMM-based features for SVM speaker verification

This paper compares several feature extraction approaches based on Gaussian Mixture Model (GMM) for Support Vector Machine (SVM) in text-independent speaker verification. Because of excellent scalability, GMM can be used to extract fixed number of typical feature vectors from various length speech data. Experiments with different GMM-based features in SVM speaker verification system were performed on the NIST’04 1side-1side database and compared with the baseline GMM-UBM.