Auto-encoder Based Channel Compensation In Speaker Recognition

As an identification technology,speaker recognition technology has been widely studied and applied.In many speaker recognition technologies,the Probabilistic Linear Discriminant Analysis(PLDA)method based on i-vector features is widely concerned,because of its good recognition result.However,the i-vector feature does not greatly distinguish between speaker information and channel information during the feature extraction process.In order to eliminate channel influence,it is necessary to compensate channel.The common channel compensation method is Linear Discriminant Analysis(LDA).LDA belongs to the linear mapping method,the ability to eliminate channel information in i-vector features is limited.Therefore,this paper proposes two new channel compensation methods based on the improvement of traditional Auto-Encoders.In this paper,the first method is named Within-class Distance Minimization Auto-Encoders(WCDM-AE).After the further improvement of WCDM-AE,the Contractive Withinclass Distance Minimization Auto-Encoders(c WCDM-AE)was obtained.Two methods have the ability of nonlinear mapping.c WCDM-AE is on the basis of WCDM-AE to add two new penalties to the loss function.The first penalty is to minimize the intra class distance;the second penalty is improved from CAE,and the I-vector and most noise are the same as the Gauss distribution,to achieve the minimum intra class distance.The channel compensation result of c WCDM-AE is better than that of WCDM-AE.The c WCDM-AE not only has the advantage of nonlinear mapping of traditional AutoEncoders,but also can use the information of category by modifying the loss function.Since i-vectors are encoded by c WCDM-AE,the features of the same speaker will be as close as possible,and using LDA will make the channel compensation better after using c WCDM-AE.It is proved by experiments that the combination of c WCDM-AE and LDA is better than channel compensation using LDA alone.In this paper,speaker verification experiment and speaker identification experiment are carried out using Voxceleb dataset.Although the performance of the c WCDM-AE+LDA method is not obvious in the experiments of speaker verification,the performance in the experiments of speaker identification is significantly improved.The c WCDM-AE+LDA achieves an absolute improvement of about 10% compared with the LDA alone.