Research On Channel And Duration Mismatch Compensation For Speaker Verification

Text-independent speaker verification technology is designed to extract the speaker’s personality information from the utterance in order to complete the verification of the speaker’s identity. Speaker verification become a research hotspot in the field of biometric identification due to its unique advantages such as convenient use and non-contact interaction. In recent years, speaker verification is gradually becoming more practical. Due to the complexity of the actual environment, it faces much problem such as the diversification of the transmission channel, the background noise pollution and so on, which led to a bad result that the performance of speaker verification system is difficult to make further improvement. The thesis focus on the speaker verification under the environmental mismatch, and discuss the methods of compensation based on total variability space and probabilistic linear discriminant analysis. For the case of duration mismatch, we proposed new method to solve the problem. The main research contents are as follows:Firstly, we discuss the extraction of Mel frequency cepstrum parameters, and the principles of Gaussian mixture model, training algorithm, the advantages and disadvantages of Gaussian mixture model for speaker verification are deeply explored. Speaker verification system based on GMM-UBM is constructed and tested by experiment.Secondly, the methods of mismatch compensation for speaker verification are deeply explored. By using factor analysis method, I-Vector of speaker from Gaussian mean super vector, then we construct speaker verification based on I-Vector. Since I-Vector contains interference information such as channel mismatch, we proposed linear discriminant analysis, intra-class covariance normalization method to compensate the I-Vector system. Experiment shows the methods can improve the performance of speaker verification system.Finally, we discuss probabilistic linear discriminant analysis method which can better modeling the speaker and interference information, and present the simplified Gaussian probabilistic linear discriminant analysis and its scoring formula. We construct speaker verification based on simplified Gaussian probabilistic linear discriminant analysis, and research on the compensation ability to I-Vector. For the duration mismatch between training utterance and test utterance, we proposed a method which can estimate duration variance information, where the duration variance information is integrated into the simplified GPLDA model. Experiments show that the proposed method can improve the performance of speaker verification.