Mismatch modeling and compensation for robust speaker verification

Abstract: In this study, primary channel mismatch scenario between enrollment and test conditions in a speaker verification task are analyzed and modeled. A novel Gaussian mixture modeling with a universal background model (GMM–UBM) frame based compensation model related to the mismatch is formulated and evaluated using National Institute of Standards and Technology (NIST) speaker recognition evaluation (SRE) 2008 data, along with a comparison to the well-known eigenchannel model. Proposed compensation method show significant improvement versus an eigenchannel model when only the supervector of the UBM is employed. Here, the supervector of the enrollment speaker model is not included for estimation of the mismatch since it is difficult to obtain the real supervector of the speaker based on the limited 5min, channel dependent speech data only. The proposed mismatch compensation model, therefore show that construction of the supervector obtained from a UBM model can more accurately describe the mismatch between enrollment and test data, resulting in effective classification performance improvement for speaker/speech applications. [ABSTRACT FROM AUTHOR]