Channel compensation technology in differential GSV-SVM speaker verification system

Channel variability is the major cause of performance degradation in text-independent speaker verification. Compensation technology in feature, model or score domain has been widely applied to baseline systems to mitigate mismatch. Newly proposed Gaussian mixture models super vector-support vector machine (GMM-SVM or GSV-SVM) baseline system has proven successful through integrating advantages of GMM and SVM. In GSV-SVM, GMM super vectors generated by stacking means of speaker models can be seen as combination of two parts: universal background model (UBM) super vector and maximum a posteriori (MAP) adaptation part. The proposed differential GSV-SVM (DGSV-SVM) removes the former to emphasize on specified speaker dependent super vector. To combat with mismatch, we put forward UBM adjustment with relative less computation, analyze the role of normalization and discuss the utility of nuisance attribute projection (NAP) in DGSV-SVM system. Experimental results on NIST SRE 06 corpus demonstrate that DGSV-SVM system with channel compensation technology has excellent performance.