VIVOLAB-UZ Speaker Diarization System for the Albayzin 2010 Evaluation Campaign

Abstract This paper describes the speaker diarization systems proposedby the VIVOLAB-UZ group for the Albayzin 2010 speaker di-arization evaluation. Our approaches combine recent improve-ments in the field of speaker segmentation in two speaker tele-phone conversations, using eigenvoice modeling, with the tra-ditional Agglomerative Hierarchical Clustering approach. Weare presenting two submissions. Our first system uses a simpleeigenvoice factor analysis model to extract a stream of speakerfactors for every recording that enable better speaker separabil-ity. The speaker factor stream is used for speaker segmenta-tion. Then, both the clusters obtained are agglomerated usingBayesian Information Criterion as distance metric, obtainingthe speaker labels. Our second submission is exactly the samesystem, but it uses Viterbi resegmentation to refine speakerchange points as a final step.Index Terms: Speaker diarization, Factor Analysis, intra-session variability, Agglomerative Hierarchical Clustering,Bayesian Information Criterion