Research Of Speaker Identification Algorithm Based On GMM And Improved LS-SVM

Speech signal contains a lot of information, such as the content of speech, speech language, gender of speaker and identity information. Speaker recognition extracts the features that indicate the speaker’s identity information, and then takes advantage of these features to identify the speaker. Addition to unforgettable, no memory, ease of use, speaker recognition had its unique advantages, such as low sensor cost, non-contact and remote verification. It had evolved into the field of biometrics technology.Least squares support vector machines speaker recognition system was a widely used method, which was essentially a modified form of the traditional support vector machine. It had the advantage of sample training speed, effectively overcome the curse of dimensionality problem and easy to implement. However, it had not the sparsity of traditional support vector machine, the system complexity increased and affected recognition efficiency. To address this issue, the Gaussian mixture model with improved least squares support vector machines was approached.Firstly, least squares support vector machine was deeply researched, and it was applied to the speaker recognition system. Because it could avoid the need for a priori knowledge, and less learning, over learning situation. Secondly, in consideration of the issues of learning capacity and noise immunity, select model parameters method was deeply researched. K-fold cross validation method was utilized to optimize parameters. Then K means primary algorithm and fast pruning algorithm were introduced, and they were combined with least squares support vector machine to make up the shortage of sparse least squares support vector machine. Finally, in consideration of the commonly used multiple classification algorithm exists indivisible areas, so that we consider to combine the fuzzy algorithm and sparse least squares support vector machine. In the experiment, vector quantization method, log-likelihood method and traditional support vector machine methods were compared. The results show that if we consider the efficiency and recognition rate, support vector machine was the best algorithm. In consideration of speech samples library in different sizes, the traditional support vector machine, least squares support vector machine and sparse least squares support vector machine were compared. The efficiency and recognition rate of sparse least squares support vector machine algorithm were both improved. Finally, compared sparse least squares support vector machine and fuzzy least squares support vector machine. The results show that fuzzy least squares support vector machine was better.