Bayesian HMM based x-vector clustering - VBx

Diarization is the task of determining the number of speakers and "who speaks when" in a recording. It is part of speech data mining. The proposed software contains a full implementation of a Bayesian approach to do speaker diarization using low-dimensional neural representation of speakers (x-vectors) in individual segments. It follows the Brno University of Technology recipe for the Second DIHARD Diarization Challenge Track 1, where BUT was the winner.It consists of computing filter-bank features, computing x-vectors, performing Agglomerative Hierarchical Clustering on x-vectors as a first step to produce an initialization, applying Variational Bayes HMM over x-vectors to produce the diarization output, and scoring the diarization output. The software is written in Python and released as open-source under Apache License.