Analysis of Speaker Recognition Systems in Realistic Scenarios of the SITW 2016 Challenge

In this paper, we summarize our efforts for the Speakers In The Wild (SITW) challenge, and we present our findings with this new dataset for speaker recognition. Apart from the standard comparison of different SRE systems, we analyze the use of diarization for dealing with audio segments containing multiple speakers, as in part of the newly introduced enrollment and test protocols, diarization is a necessary system component. Our state-of-the-art systems used in this work utilize both cepstral and DNN-based bottleneck features and are based on i-vectors followed by Probabilistic Linear Discriminant Analysis (PLDA) classifier and logistic regression calibration/fusion. We present both narrow-band (8 kHz) and wide-band (16 kHz) systems together with their fusions.