The Study Of Speaker Recognition System Based On GMM Algorithm

Speaker recognition technology is a kind of biological certification technology, by analyzing the voice of the speaker’s personality characteristics to automatically identify the speaker’s identity. Speaker recognition technology with its unique economic advantages and accuracy, is widely used in people’s daily life and work, and brought a lot of convenience. Therefore, the study of a high recognition rate, and robustness of speaker recognition method is the direction for the majority of researchers.This study is focused on the speaker recognition system based on Gaussian mixture model. After some of the parameters in the system and the model test and comparison, proposes the corresponding improvement to the system, to achieve the purpose of improve the system recognition rate. In this paper, the system construction, system performance aspects of research and system improvements were studied experiments.First, the use of Matlab to build a Gaussian mixture model based speaker recognition system, specifically including selection and read speech database, training and recognition of several modules of the speech signal preprocessing, feature extraction of speech signal parameters and Gaussian mixture model.Research on system performance is to study the speaker speech feature extraction and model parameters affect system performance. Test shows that the performance characteristics of different parameters of the test voice different lengths and different frame sizes and the type of test speech model order will affect the system. MFCC parameter better recognition rate LPC and LPCC, test speech will increase the length of the proper extension system, Gaussian mixture model under the same number of high-end frame length is superior to low-level model, Gaussian mixture model for a different order, the system when the voice of the Department of optimal frame size can also vary.In this paper, the use of GMM speaker model training the model, but also the use of the BP model to be trained to recognize the speaker’s voice. For BP model uses MFCC feature extraction parameters were obtained recognition rate of different training set size. The results found that with the increase of speaker training speech, the recognition rate will decline. Delta introduction to this problem and proposes a method characterized in the second feature extraction, and the MFCC feature LPCC processing is performed again, and then filter the free combination of the characteristics of the most efficient test system, the test proved that the improved characteristics of speaker identification when increasing the number of the effect is not significantly reduced.