Speaker Verification Based On Factor Analysis

This thesis focus on the issue of channel mismatch in speaker verification channel mismatch is one of the most factors impacting System's Performance, if we want the system to have better robustness, that channel compensation should be used to compensate the mismatch information. The GMM-UBM is the state-of-the-art system in speaker verification, so this thesis focus on the method that used in this system's channel compensation. We use the factor analysis to estimate the channel subspace and speaker subspace in the total space, so can separate the speaker information and channel information.1) the channel mismatch information will affect the system's performance ,so we compare several channel compensation techniques , find the Advantages and disadvantages within them, in the GMM-UBM speaker system, finally using the factor analysis methods for our speaker system.2)Taking Computation, Complexity and speech data into account, we simply the joint factor analysis method, when modeling speaker information, we use the MAP to estimate speaker information, so it doesn't need additional speech data to model speaker space, directly use the UBM's Covariance.3) The most important question is how should we use the factor analysis for compensating the mismatch between the training speech and test speech. We introduce two Compensation algorithmn that used in the Feature Domain and Model Domain.Feature Domain algorithm:We use FA to estimate the test speech's channel factor, then map the feature, and we have the new features which do not contain the channel information. And then use it to test GMM-UBM system, where GMM is the speaker model ,and it also doesn't contain channel information.Model Domain algorithm:This algorithm need to estimate a new LLR for GMM-UBM system, we introduce several methods to estimate LLR, Taking Computation and performance into accout, we use the linear Scoring methods. 4)The latest state-of-the-art Joint factor analysis (JFA) speaker verification system's scoring methods need large and complex calculations,we will directly use speaker factor for scoring。First,both on training and testing we need to estimate the speaker factor by factor analysis,then use them for scoring,comparing with other systems like SVM and other JFA GMM-UBM system,we do not need other process for training enrollment and less storage。Experiments based on NIST2008 male speaker corpus show that use cosine distance scoring make the process faster and less complex than other FA scoring.