Comparison Studies of Speaker Modeling Techniques in Speaker Verification System

Abstract

In this paper a brief comparison studies on the performance of different speaker modeling techniques in robust and reliable speaker verification (SV) system has been discussed. In text-independent speaker verification, lots of states of art speaker modeling techniques have been developed in different scenarios to upgrade its performance. The performance of SV system is not only depended on the fusion of different feature vectors but also it is highly depended upon the fusion of various speaker modeling techniques. In this work, an automatic SV system has been developed using the Mel-Frequency Cepstral Coefficients (MFCC) combined with the Prosodic feature vectors. The baseline of the SV system has been trained with speaker modeling techniques separately and fusions namely Vector Quantization (VQ), Gaussian Mixture Model (GMM), GMM-Universal Background Model (GMM-UBM), Support Vector Machine (SVM) and Joint Factor Analysis (JFA) to analyze its performances. The results reported here, have been evaluated using the multilingual speech database, namely Arunachali Language Speech Database (ALS-DB). From the experimental point of view we observe that the best performance of SV system shows by JFA with GMM-UBM modeling technique with its EER value of 4.76% and MinDCF value of 0.0872. Comparing with other modeling techniques VQ shows its poor performance with its EER value of 11.08% and MinDCF value of 0.2010. SVM shows of approximately 2.8% improvement of verification rate with comparison to that of GMM-UBM. Here, finally, we conclude that the fusions of both generative and discriminative models highly improve the performance of SV system.