Evaluating speaker recognition systems under challenging conditions with mimicked voices of target speakers as non-target trials is very important in making decisions of its effectiveness and deployability in critical applications. The lack of a standard database is one of the bottlenecks in pursuing research in this direction. In this paper, we present the details of a new multilingual speaker recognition evaluation database, Amrita SRE database, in which the impersonators mimic the voices of target speakers. Our database consists of 115 target speakers and 76 impersonators speaking in six different languages: English, Hindi, Malayalam, Kannada, Tamil or Telugu, makes a total of 815 target speaker models, 6994 target trials and 3976 non-target trials. We evaluate the performance of different state-of-the-art speaker recognition systems using the normal target and nontarget trials and compared its performance with non-target trials replaced with mimicked voices. For all the systems, an average performance deterioration of 10% absolute in equal error rate (EER) was observed when tested with mimicked non-target utterances.
K. K. George, Dr. Santhosh Kumar C., Sreekumar, K. T., K Das, A., Thottupattu, A. J., Kumar, M. S., and Ramachandran, K. I., “Amrita SRE Database: A Database for Evaluating Speaker Recognition Systems with Mimicked Speech”, 2015.