a-DCF: an architecture agnostic metric with application to spoofing-robust speaker verification

The tandem approach is characteristic of Standard metrics can be applied to evaluate the performance of the majority of related work, including studies involving other isolated spoofing detection solutions and others have been proposed biometric traits [10, 11]. to support their evaluation when they are combined with Standard metrics developed for the evaluation of speaker speaker detection. These either have well-known deficiencies or detectors can also be applied to the evaluation of spoof detectors, restrict the architectural approach to combine speaker and spoof also known as countermeasures (CMs); they are both binary detectors. In this paper, we propose an architecture-agnostic classifiers. Alternative metrics proposed in recent years also detection cost function (a-DCF). A generalisation of the original support the evaluation of speaker and spoof detectors when DCF used widely for the assessment of automatic speaker combined [12, 13]. While the combination of speaker and spoof verification (ASV), the a-DCF is designed for the evaluation detectors still constitutes a single, binary classifier with the very of spoofing-robust ASV. Like the DCF, the a-DCF reflects the same original task of accepting bonafide target trials and rejecting cost of decisions in a Bayes risk sense, with explicitly defined anything else, the consideration of spoofing complicates class priors and detection cost model.