Research On Speaker Verification Based On Latent Information Learning

Speaker verification is a biometric technology that can effectively identify the speaker according to the speaker's voice characteristics.With the development of science and technology,it has been made great progress in speaker verification,and it has gradually developed in practical applications.How to condense the accurate speaker representation from a large number of data has become the most noteworthy issue in speaker verification.Therefore,it is necessary to further investigate speaker representations that can directly represent the identity characteristics of different speakers.However,the current methods cannot make full use of the latent information represented by speaker feature space and speaker representations.Based on the above analysis,the dissertation focuses on three issues:(1)how to make full use of the prior information of the speaker feature space;(2)how to reduce the influence of speaker feature space from class imbalance;(3)how to mine the correlation information between the speaker representations.Ultimately,it will achieve the goal of improving the performance of the speaker verification system.The main contents of the dissertation are summarized as follows:(1)In the aspect of making full use of the prior information of speaker feature space,a speaker verification method based on Bayesian principal component analysis(BPCA)is proposed.This method introduces the BPCA into the learning of total variability space(TVS)and supplements the missing information of the development data with a-priori distribution assumption.It can highlight the role of effective dimensions in the TVS and weaken the influence of invalid dimensions under the constraint of a-priori information.Ultimately,it will increase the effective information contained in the TVS and enhance its ability to represent the speaker information.Experimental results show that when the development data is insufficient,compared with the TVS learning methods,the proposed method effectively supplements speaker information and improves system performance.(2)In the aspect of reducing the influence of speaker feature space from class imbalance,a speaker verification method based on class contribution balancing is proposed.In order to balance the contribution of different classes to feature space,the balance factor based on the multi-classification cross entropy loss is introduced in this method.Accordingly,the proposed method can focus on small-scale samples and ensures the attention of large-scale samples.And it can mine the latent speaker information fairly.Experimental results show that compared with the typical multi-classification loss function,the proposed method can effectively alleviate the impact of class imbalance,thereby improving the performance of the system.(3)In the aspect of mining the correlation information between speaker representations,a speaker verification method based on maximizing the divergence of deep information is proposed.The information divergence that can calculate the similarity between distributions is introduced into the objective function.The proposed method maximizes the statistical correlation between representations.And it can control the neural network to mine the latent compatibility information between similar representations and improve the difference of heterogeneous features in the feature space.Meanwhile,the relations of the deep feature space can be described more accurately and the discriminability of the speaker representation can be improved effectively.Experimental results show that compared with other deep learning methods,the proposed method significantly improves the performance of speaker verification system.