PLDA-based Speaker Verification in Multi-Enrollment Scenario using Expected Vector Approach

Multi-Enrollment scoring scenario, where multiple utterances are available for an enrollment speaker, is one of the less explored problems in the Probabilistic Linear Discriminant Analysis (PLDA) scoring literature. Since the closed-form PLDA scoring formula for multi-enrollment scenario is impractical, alternate heuristic approaches are widely used for such scenarios in both i-vector and x-vector based speaker verification systems. In this paper, we describe an Expected Vector approach to obtain a vector from multiple enrollment utterances. Expected Vector approach uses a trained PLDA model to compute the expected class center given a set of vectors for that particular PLDA model. By using such an approach, a more meaningful class center representation can be obtained. This vector can be used to score a trial using two-vector scoring formula for a given PLDA model. We compare the performance of the proposed approach with various heuristic approaches and show that it provides significant improvements in terms of Equal Error Rate (EER) and minimum Detection Cost Function (minDCF). We show our results on x-vector system trained on Voxceleb dataset with various implementations of PLDA and trials designed on Voxceleb and Librispeech dataset.