Speaker Verification Technology Based On Tensor Structure

Speaker verification technology is an important research task in speech signal processing,and there's an important means to biometric authentication technology.The major work is to check if the test speech belong to the target speaker by comparing their model that be established based on the extraction of the target speaker's training speech and the acoustic characteristics of the test speech.In order to improve the performance of speaker verification system in noisy environment,this paper proposes an improved MFCC feature parameter extraction method based on tensor structure.The technology constructs a logarithmic filter bank energy of different speakers into a third-order tensor,and obtains a projection matrix by tensor PCA analysis with non-negative constraints.And we projects the log filter bank energy with the matrix before the DCT transform.We can obtained T-MFCC feature parameter,a new and improved feature.The experimental results show that compared with the traditional MFCC,the T-MFCC is less affected by low SNR.And the T-MFCC-based speaker verification system has a lower error rate in the same noisy environment.The reason is that the tensor PCA method better preserves the original contact data between the different speakers while completing the dimensionality reduction and de-noising tasks.Instead of ordinary methods that are prone to losing useful information,the tensor model can not only preserve the intrinsic relationship between different speakers,but also dig out the useful information hidden between different sessions of the same speaker.This paper proposes a speaker modeling method based on non-negative tensor decomposition technique for classification and recognition.We constructed the feature tensor with the feature extracted from the training speech,for each speaker.And the dimension of tensor is feature?frame?session.We used the ALS to obtain the non-negative tensor decomposition of feature tensor,and got the nuclear tensor Gi and the factor matrix.Finally,we projected the features of the test speech into the subspace corresponding to the factor matrix.We get the result of verification by comparing it with Gi.Experiments show that the speaker model of tensor structure has higher accuracy than other traditional modeling methods for speaker verification.In summary,the work of this paper is to use the data analysis method of tensor to extract the robust feature of T-MFCC.And we constructed the speaker model with the non-negative tensor decomposition technique to achieve the aim of improving the accuracy of system identification.