Speaker recognition in two-speaker data: recent results from Dragon Systems

Abstract

NIST has recently introduced two-speaker tasks in its Speaker Recognition evaluation scheme. We discuss Dragon Systems' recent explorations of these tasks using a GMM-based speaker ID system. We present baseline results submitted for the NIST two-speaker evaluation in April 1999. We also report on sonie "cheating" experiments and investigations of alternative segmentation and channel normalization techniques, which illuniinate some challenges specific t,o the two-speaker environment.