The performance of speaker recognition, as measured by NIST in its periodic evaluation programs of speaker detection technology, is affected by the predictive accuracy of system decisions as well as the inherent discriminating power of the underlying algorithms. This paper analyses the score calibration accuracy of a few systems that exhibited good performance in the NIST 2010 speaker recognition evaluation (SRE10) and proposes a modified cost function that better represents the accuracy of score calibration in the low false alarm decision region, which is the region of interest for intelligence applications.

Index Terms: speaker recognition, speaker detection, score calibration