Policy analysis framework for conversational biometrics

Modern speaker verification systems rely on a variety of input sources in making a decision on the validity of an identity claim. The meaning of the evidence these sources produce must be reconciled if robust decisions are to be made. In the case of Conversational Biometrics (CB), for example, this is typically accomplished via the specification of a verification policy implemented as a finite state machine. This paper presents a framework for the analysis of such complex systems guided by a policy finite state machine. The Receiver Operating Curve (ROC) associated with the acoustic speaker recognition task is transformed into a multi-dimensional Receiver OperatingMap (ROM), which results from a probabilistic analysis of the policy state machine. A DCFMap can be similarly generated and we show results indicating that optimization over this surface is an appropriate way to set thresholds.