Speaker Verification on Summed-Channel Conditions with Confidence Measures

Abstract

Abstract. This paper addresses the problem of speaker verification in two speaker conversations, proposing a set of confidence measures to assess the quality of a given speaker segmentation. We study how these measures can be used to estimate the performance of a state-of-the-art speaker verification system, the I3A submission for the core-summed condition in the NIST SRE 2010. We present a Factor Analysis based speaker segmentation system, along with three confidence measures that are fused to obtain a single measure that we show to constitute a good estimation of the segmentation accuracy, when evaluated on the summed-channel telephone data of the NIST SRE 2008. Finally we present speaker verification results obtained with the I3A submission for the NIST SRE 2010 on several conditions of this evaluation, involving summed-channel. We show that the confidence measure also predicts the performance of a state-of-the art speaker verification system when it faces two speaker conversations.