Speaker Recognition on Single- and Multispeaker Data

We discuss Dragon Systems’ approach to the NIST Speaker Recognition tasks. For the one-speaker task, we employ a combination of methods: a basic GMM system and two LVCSR-based systems, one using standard mixture models and the other using nonparametric techniques. We discuss some explorations of the recently introduced two-speaker tasks based on the GMM system alone. “Cheating” tests using NIST-supplied keys lead us to some improvements in channel normalization, and illuminate the roles that speaker segmentation and segment selection play in these tasks.