Speaker Diarization based on Bayesian HMM with Eigenvoice Priors

Nowadays, most speaker diarization methods address thetask in two steps: segmentation of the input conversation into(preferably) speaker homogeneous segments, and clustering.Generally, different models and techniques are used for the twosteps. In this paper we present a very elegant approach where astraightforward and efficient Variational Bayes (VB) inferencein a single probabilistic model addresses the complete SD problem.Our model is a Bayesian Hidden Markov Model, in whichstates represent speaker specific distributions and transitions betweenstates represent speaker turns. As in the ivector or JFAmodels, speaker distributions are modeled by GMMs with parametersconstrained by eigenvoice priors. This allows to robustlyestimate the speaker models from very short speech segments.The model, which was released as open source codeand has already been used by several labs, is fully describedfor the first time in this paper. We present results and the systemis compared and combined with other state-of-the-art approaches.The model provides the best results reported so faron the CALLHOME dataset.