Speaker Recognition Based On Sparse Representation And Channel Compensation

First,second order difference of MFCC feature parameters are used to characterize the speaker’s speech features.Then using the classical GMM-UBM model to train templates.On this basis,the I-vector algorithm is added.Compared with the method of joint factor analysis,I-vector only needs to train the global difference space,with expressing the speaker and channel related information.Each speaker’s speech feature is represented by I-vector vector.The application of probabilistic linear discriminant analysis(PLDA)in the I-vector space training would reduce the dimensionality and compensate the channel,which can improve the performance of the system to some extent.Then,on the basis of i-vector,sparse representation and the sparse representation classifier based on i-vector are introduced.By constructing a i-vector Dictionary of training speech extractions,the test speech extraction i-vector is represented as a linear combination of dictionary atoms.The signal is reconstructed according to the sparse representation coefficient,and the classification of the speech signal is determined by the residual error between the reconstructed signal and the original signal.The higher the GMM mixing degree is,the higher the recognition rate of the system is,but the complexity of the system computation increases as well.For this reason,PPCA algorithm is introduced to reduce dimension.Compared with the traditional principal component analysis(PCA),PPCA considers the probability distribution of components,principal components selected,overcoming the PCA simply select several characteristics of maximum value of the corresponding feature vectors,which can not guarantee the selected principal components must be the biggest contribution of feature vector.Through the PPCA correlation,optimal dimensionality reduction and noise reduction experiments,and PPCA and PCA contrast test,we can see that the PPCA dimension reduction can maintain the recognition rate,or even slight growth.Meanwhile,the dimensionality reduction performance of PPCA is better than that of PCA.Finally,this paper makes a detailed summary of the research,summarizing the contribution and shortcomings of this paper.At the same time,the future development of speaker recognition is prospected.