Font Size: a A A

Research Of Robust Speaker Verification Method Based On Factor Analysis

Posted on:2010-07-05Degree:MasterType:Thesis
Country:ChinaCandidate:D H WuFull Text:PDF
GTID:2178360302959545Subject:Circuits and Systems
Abstract/Summary:PDF Full Text Request
With growing requirements in practical, the speaker verification systems are not only required to obtain good results in the laboratory conditions, but also need to have better robustness in a complex background. Channel has been the factor which leads biggest impact of speaker verification system, because of the complexity of communication lines, as well as the diversity of the microphone. While training speech and testing speech channels mismatch, it leads to a sharp decline in verification performance.Firstly, this thesis analyzes speaker verification system based on the probability statistical model and identify model. Secondly, it posts a method, which is the combination of the probability statistical model and identify model. Then, we analyzed the current channel mismatch compensation methods, researched the channel compensation based on factor analysis. The main research contents are as follows:1. Aim at improving the distinction of probability statistical models system, and dealing with the problems that lack of speaker personalized information description while using distinguish identify model, we present a combination of the GMM-SVM speaker verification method. The use of GMM as the front-end characteristics of transformation and clustering, and then based on support vector machine to distinguish training the speaker model, which combined the merits of the probability statistical model and the distinguish identify model. Comparative experiment based on NIST database shows that distinguish training can significantly improve system performance of the probability statistical model.2. In order to resolve the channel mismatch problem while speaker verification systems are in complexity environment, we give a simplified algorithm of factor analysis based on the framework of GMM_UBM system. The combination between correlation coefficient of MAP and factor analysis has greatly reduced the complexity. After channel space estimated, the feature mapping methods used as front-end to remove channel characteristic parameter information. To retain the integrity of the framework of the GMM_UBM, it limited the factor analysis only to front-end parameters, which reduces the computational complexity significantly.3. In order to integrate the merits of factor analysis and distinguish identify model, this article proposed a speaker verification method which combines factor analysis and support vector machine. Removed channel information, the GMM super vector as SVM speaker verification system input parameters, which combines the advantages of identifying model. Experiments proof the performance of this system been further rose than the system in GMM_UBM framework based on factor analysis.
Keywords/Search Tags:speaker verification, channel mismatch, factor analysis, Gaussian mixture models, support vector machine
PDF Full Text Request
Related items