Novel Architectures for Unsupervised Information Bottleneck Based Speaker Diarization of Meetings

Speaker diarization is an important problem that topical, and especially useful as a preprocessor for conversational speech related applications. The objective of this article two-fold: (i) segment initialization by uniformly distributing speaker information across the initial segments, (ii) incorporating discriminative features within unsupervised framework. In first part work, varying length technique Information Bottleneck (IB) based system using phoneme rate side proposed. This distributes segments provides better starting point IB clustering. second we present Two-Pass (TPIB) incorporates during process diarization. TPIB has shown improvement over baseline system. During pass system, coarse segmentation performed alignments obtained are used to generate shallow feed-forward neural network linear discriminant analysis. in obtain final boundaries. paper, variable combined with leverages advantages results additional performance. An evaluation on standard meeting datasets shows significant absolute 3.9% 4.7% NIST AMI datasets, respectively.