Research On Speaker Confirmation Based On SMFCC Feature And Factor Analysis

Speaker recognition is the technology that identify the speaker’s identity by speaker’s voice. As an important identify technology, the speaker recognition has been applied in the fields such as national security, public security and authentication, etc., in recent years. At the same time, there are still some problems to be solved, such as channel diversification identification, noise impact on recognition performance, security anti-counterfeiting performance promotion, which is related to extraction speech signal feature and speaker recognition algorithm. Therefore, it is very meaningful to study the speaker recognition. This paper mainly focuses on the speech feature extraction and recognition algorithm of the speaker verification, and mainly achievements are as follows:A novel feature extraction method called SMFCC for speech signal is proposed. We study the time domain and frequency domain characteristics of the speech signal, and made an exploratory research that use the S-transform to process the speech signal. Combined with the existing parameters of speech signal, we propose a novel speech feature extraction method. At the same time, we try to use Singular Value Decomposition (SVD) to reduce the noise of speech signal which exists the bandwidth of random noise, and on this basis, an adaptive Singular Value Decomposition (ASVD) filtering algorithm is proposed. In order to verify the effectiveness of the novel method, we compare the novel feature and some of existing features such as LPCC, MFCC and LPMFCC, etc. in speaker recognition system, the results show that the performance has obvious improvement.A better performance model for speaker recognition is obtained, which is Gaussian Mixture Super Vector-Support Vector Machine (GSV-SVM). We put GMM super vector which is on the basis of the GMM-UBM into SVM, and the GSV-SVM is obtained. On this basis, we study the common kernel functions of SVM, and obtain the optimal kernel function of GSV-SVM by comparing in the speaker recognition system. The GSV-SVM model based on the optimal kernel function is compared with GMM-UBM, results show that the performance improved significantly.A total variation space estimation algorithm called Universal Background-Joint Estimation (UB-JE) based on Total Variation Factor Analysis method (i-vector) is proposed. By studying the factor analysis method, it is concluded that the estimates of total variation factor space in the i-vector model plays a key and basic role. We put forward a A total variation space estimation algorithm called UB-JE by starting the Factor Analysis methods. Firstly, the total variation space Universal Background algorithm (UB) is proposed according to the thought of GMM-UBM. Secondly, we propose a total variation space Joint Estimation algorithm (JE) according to the Factor Analysis theory and related literature. Finally, we combine the two algorithms to obtain the UB-JE. In order to verify the effectiveness of the algorithm, which is integrated into the i-vector, and compared with the conventional i-vector algorithms, results show that performance has obvious improvement.All methods proposed in the paper, which are the new feature extraction algorithm, the new total variation factor space estimation algorithm and the SVM based on the optimal kernel function are integrated into a speaker recognition system with better performance.