Comparative analysis of two different system's framework for text dependent speaker verification

摘要

Speaker verification is among the widely used biometrics which generally offers more secure authentication for user access compared to regular passwords. Speaker verification is the process of automatically authenticating the identity of the speaker in order to protect the resources by controlling its access. Usually identity is claimed by presenting a unique personal possession, here in this case a sample of the user's voice. In this paper we present two different framework for text dependent speaker verification, one using Hard threshold based system and the other using a Cohort Based System. The baseline system used employs MFCC and DTW for their verification purposes. A database of 30 Speakers collected over practical noisy environment was used for testing and validating the modules. Experimental results shows that the Cohort based speaker verification system achieves good performance compared to a hard threshold system on a text constrained speaker verification task. Finally the combined system based on the normalized performance score of individual techniques outperforms the stand alone system and increases the performance to 85.61% for practical noisy conditions respectively. The approaches are described and detailed experimental results and analysis are presented and discussed.