Research On Mismatch Compensation For Speaker Verification

The issue of mismatch between train and test speech is the key challenge inthe current text-independent speaker recognition (especially for telephone/mobilespeech under the varying environment in NIST SRE). This mismatch can begrouped into two catalogs: content and channel variability, such as the di?erenttrain and test speech content, handset type, channel transmission e?ects andnearby sources of noise. This list does not even include many of the potentialsources of mismatch introduced by the speaker themselves. How to solve themismatch issue now is becoming the focus topic in the teld of speaker veri?cation.This thesis focuses on the issues of mismatch in text-independent speakerverifcation which the telephone speech data is recorded under the complex envi-ronment. Aiming at the mismatch compensation in speaker veri?cation, in thisthesis, we mainly study in the following aspects: more accuracy mismatch com-pensation, improving the speed of the implementation method and applying themethod in discriminate speaker recognition system.We frstly review the details of the current mismatch compensation methods,and discuss How the UBM plays an important role in estimating target modeland test segment score in the GMM-UBM framework recognition system. Wealso introduce and compare several techniques which are used to compensate thesession variability, and find these techniques fail to meet the goal stated above fordiferent reasons.In an attempt to overcome the above mentioned problem, we present a mis-match compensation approach based on factor analysis algorithm, and describehow this approach addresses the issue of mismatch in GMM-UBM-based speakerverifcation by explicitly modeling session variability in both the training and test-ing procedures and learning from the mismatch encountered. By directly model- ing the mismatch between sessions in a constrained subspace of the GMM speakermodel means, the proposed technique replaces the discrete categorization of tech-niques with a continuous vector-valued representation of the session conditions.A major strength of this approach is that the training methods used also removethe need for labeling the training data for particular conditions. This approachwas successfully used in our lab's speaker veri?cation system in NIST SRE 2008.The major deficiency of factor analysis algorithm is that huge computationalcost is required for each frame of speech is computed EM statistics with all mix-tures of UBM. This thesis proposes a fast mismatch compensation algorithm basedon CUBM selective model for speaker verification and demonstrates how it suc-cessfully implemented in NIST SRE. We also propose a novel Top-N selectivestrategy to select mixtures of UBM for improving the accuracy of selecting pro-cess. We tested the novel algorithm on the NIST SRE 2006 evaluation set. Ourpresented system obtains almost the same performance as the factor analysis base-line system while at the same time reducing the real-time factor by a factor ofeight.We also present an approach to combine the compensation technique anddiscriminate recognition system together. This novel approach uses the GMMspeaker supervector produced by Factor Analysis target speaker model as inputfor a SVM based on classic linear kernel between two supervectors. Accordingto our experiments on NIST SRE 2008, the mismatch compensation techniquesigni?cantly improve the performance of discriminate system, our proposed systemobtains 24% relative improvement in EER compared to the baseline discriminatesystem.This thesis was supported by State Scholarship (No.2009634072) provided byChina Scholarship Council and the innovation foundation for graduate student ofUSTC (No.KD2008056).