Speaker verification using adapted Gaussian mixture models

In this paper we describe the major elements of MIT Lincoln Laboratory's Ga ussian mixture model (GMM)-based speaker verification system used successfu lly in several NIST Speaker Recognition Evaluations (SREs). The system is b uilt around the likelihood ratio test for verification, using simple but ef fective GMMs for likelihood functions, a universal background model (UBM) f or alternative speaker representation, and a form of Bayesian adaptation to derive speaker models from the UBM. The development and use of a handset d etector and score normalization to greatly improve verification performance is also described and discussed. Finally representative performance benchm arks and system behavior experiments on NIST SRE corpora are presented. (C) 2000 Academic Press.