Intra-session variability compensation and a hypothesis generation and selection strategy for speaker segmentation