Speaker Diarization Based on Gmm Supervectors and Unsupervised Intra-speaker Variability Modeling

This paper presents a novel framework for speaker diarization. Audio is parameterized by a sequence of GMM-supervectors representing overlapping short segments of speech. Session dependent intra-session intra-speaker variability is estimated online in an unsupervised manner, and is removed from the supervectors using Nuisance Attribute Projection (NAP) The supervectors are then projected using principal component analysis (PCA) into a one-dimensional space (for the two speaker case). Viterbi decoding is used to find an optimal segmentation using the projected supervectors as estimates for log-likelihood ratios. Finally, GMMs are adapted for each assumed speaker and Viterbi re-segmentation is used for producing a final diarization.