Based On The Sparse Representation Of Speaker Recognition Research

With the continuous development of modern biological recognition technology, speaker recognition technology has been more and more attention of experts and scholars. Through the analysis of the voice of the speaker information, so as to effectively identify the identity of the speaker, speaker recognition technology has been widely applied because of its simple and efficient features, such as network security, judicial authentication and the national information security and other fields.In the traditional method of speaker recognition, GMM- UBM using general background template adaptive speaker model, reflects the personality characteristics, but the amount of calculation is too large, distinguish between ability is insufficient, and popular SVM classifier using nonlinear kernel function to classification, although the recognition performance is improved, but it is overly dependent on the distribution of sample data. Sparse representation theory can reflect the characteristics of the signal using the minimum number of atoms, so as to further improve the system to distinguish. So the in-depth analysis of sparse representation, on the basis of the principle and the dictionary structure method was proposed based on sparse representation of speaker recognition, at the same time Fisher discriminant dictionary learning algorithm was applied to speaker recognition, specific work is as follows:(1) the research learning sparse representation theory, analysis of speech signal sparse sex, to explore the feasibility of speech signal sparse sex said, combination with the characteristics of speech signal of reconfigurable,speaker recognition method based on sparse representation i- SRC, using the current mainstream of I-vector modeling speaker model, vector structure random to use the average of the super dictionary, again through the reconstruction error identification, experiments have shown that GMM- SRC recognition rate is improved.(2) in order to further improve the efficiency of recognition, based on i- SRC recognition method, study discusses the construction method of a dictionary, introduces a dictionary with discriminant method of study, namely, Fisher discriminant dictionary learning algorithm, using the algorithm is obtained with a distinct training of structured dictionary, the dictionary of atoms and there is a corresponding relationship category labels, this is a good way to use refactoring information classifying test voice. Speech database on the NIST experiments also show that the introduction of new dictionary algorithm improved the recognition method for error rate declined, and the recognition efficiency is also improved.