Challenging margin-based speaker embedding extractors by using the variational information bottleneck

Speaker embedding extractors are typically trained using a
classification loss over the training speakers. During the last
few years, the standard softmax/cross-entropy loss has been
replaced by the margin-based losses, yielding significant im-
provements in speaker recognition accuracy. Motivated by
the fact that the margin merely reduces the logit of the target
speaker during training, we consider a probabilistic framework
that has a similar effect. The variational information bottle-
neck provides a principled mechanism for making deterministic
nodes stochastic, resulting in an implicit reduction of the pos-
terior of the target speaker. We experiment with a wide range
of speaker recognition benchmarks and scoring methods and re-
port competitive results to those obtained with the state-of-the-
art Additive Angular Margin loss.