Analysis of Critical Metadata Factors for the Calibration of Speaker Recognition Systems

LUCIANA FERRER; AARON LAWSON; MAHESH KUMAR NANDWANA; DIEGO CASTÁN; MITCHELL MCLAREN

Graz

Congreso; Interspeech 2019; 2019

ISCA

In this paper, we analyze and assess the impact of critical meta-data factors on the calibration performance of speaker recog-nition systems. In particular, we study the effect of duration,distance, language, and gender by using a variety of datasetsand systematically varying the conditions in the evaluation andcalibration sets. For all experiments, the system is based on i-vectors and a probabilistic linear discriminant analysis (PLDA)back-end and linear calibration. We measure system perfor-mance in terms of calibration loss. Our experiments reveal (i) alarge degradation when the duration used for calibration is sig-nificantly different from that in the evaluation set; (ii) no signif-icant degradation when a different gender is used for calibrationthan for evaluation; (iii) a large degradation when microphonedistance is significantly different between the sets; and (iv) asmall loss for closely related languages and languages withshared vocabulary. This analysis will be beneficial in the de-velopment of speaker recognition systems for use in unseen en-vironments and for forensic speaker recognition analysts whenselecting relevant population data.