Score normalization for VQ-UBM based text-independent speaker verification

This paper presents score normalization for recently proposed modeling technique, vector quantization - universal background model (VQ-UBM) based speaker verification of cellular data. Test-normalization (TNorm) which is the most widely used score normalization technique, is evaluated for VQ-UBM based speaker verification. Experimental results using NIST 2002 Speaker Recognition Evaluation (SRE) (one-speaker detection task) show that score normalization improves the verification performance and VQ-UBM provides better recognition accuracy than support vector machines - generalized linear discriminant sequence kernel (SVM-GLDS), which is one of the state-of-the-art modeling techniques for speaker verification, in terms of both, Equal Error Rate (EER) and Minimun Detection Cost Function (MinDCF).