A Robust SVM/GMM Classifier for Speaker Verification

One of the basic problems in the speaker verification applications is presence of environmental noise. State-of-art speaker verification models based on Support Vector Machine (SVM) show significant vulnerability to high noise level. This paper presents a SVM/GMM classifier for text independent speaker verification which shows additional robustness. Two techniques for training GMM models are applied, providing different results depending on the values of environmental noise. The recognition phase was tested with Serbian speakers at different Signal-to-Noise Ratio (SNR).