A Hybrid Generative-Discriminative Approach to Speaker Diarization

摘要

In this paper we present a sound probabilistic approach to speaker diarization. We use a hybrid framework where a distribution over the number of speakers at each point of a multimodal stream is estimated with a discriminative model. The output of this process is used as input in a generative model that can adapt to a novel test set and perform high accuracy speaker diarization. We manage to deal efficiently with the less common, and therefore harder, segments like silence and multiple speaker parts in a principled probabilistic manner.