Interchangeability of Calibration Audio Datasets for Forensic Automatic Speaker Recognition

When employing automatic speaker recognition in forensic voice comparison, the practitioner must choose a set of calibration audio data that is representative of the case audio in order to train a score-to-LLR function. The decision of which data constitutes representative data and which audio and speaker conditions should be considered is ultimately a subjective judgment by the practitioner. In this work, two metrics and a graph that show interchangeability of potential calibration datasets are proposed. They are designed to help inform the decision of representativeness by the practitioner by showing the effect of some audio condition on the score-to-LLR function and the resulting LLRs.