Analysis of back-end classifiers for speaker verification with DNN embeddings

The rapid evolution of DNN-based speaker embeddings has brought significant improvements in the field of speaker recognition. The increased focus on DNN architectures, however, has been accompanied by a reduced interest and a significant simplification of the embedding classification stage, due in part to the poor performance of Probabilistic Linear Discriminant Analysis (PLDA) and in part to the success of cosine scoring for comparing speaker embeddings. As a result, PLDA has been gradually replaced by cosine scoring. In this work, we critically evaluate the effectiveness of cosine distance with respect to alternative generative and discriminative classification methods. We compare cosine distance with several alternative classifiers over a large set of publicly available benchmarks with different supervised DNN embedding extractors, showing that substantial gains can be obtained by more complex back-end models, whereas over-reliance on cosine scoring may, in some cases, even lead to misleading conclusions on the relative performance of different DNN architectures.