Research Of Speaker Recognition System Based On Mixed Festure Parameters And GMM-UBM

With the rapid development of Internet technology and information technology, the technology of speaker recognition with voice comes into being. Because of its non-invasive and the most natural and intuitive way for users, the speaker recognition technology becomes one of the most acceptable ways of biometric authentication, and it is widely used in all areas of society, such as the judicial investigation, e-commerce, finance and so on.Although speaker recognition technology has achieved satisfactory results in theory, but there are still some problems in practical application that need us to further study. In this paper, from the overall framework of speaker recognition system, based on the analysis of existing speaker recognition technology, research on the front-end processing of the speech signal, feature extraction, pattern matching and other issues. The specific work and innovations are as follows:(1)Due to the accuracy of the voice activity detection affecting the recognition performance of the entire system, and the poor anti-noise performance of short-time TEO energy, In order to solve the problem, this paper proposes a new voice detection algorithm in noisy environment. On the basis of short-time TEO energy voice detection, adding the judging section Mel Cepstral distance, we use two-stage decision mechanism with accurate judgment after rough judgment. Compared with the traditional double thresholds and spectral entropy in the environment of different noise and different signal to noise ratio, the experiment results show that this designed algorithm has better capability in low SNR and improves the accuracy of the detection system without increasing the computational complexity.(2)The selection of characteristic parameters affect the speaker recognition system, in order to extracting the characteristic parameters that reflecting the speaker’s personality characteristics. In this paper, the Mel Frequency Cepstral(MFCC) that reflecting individual voice characteristics and short-time TEO energy that reflecting time-domain characteristics of the speech signal mixed feature parameters are adopted to the speaker recognition system, this method increases the effective dimension of features to improve the shortage of the characteristics sample. And using correlation distance Fisher criterion select feature vector that making more contribution and regroup them, then we get better mixing parameters indicates the speaker’s voice characteristics.(3)The paper analysis the main technology of the speaker recognition system and build speaker recognition model based on the GMM-UBM. Through training all speakers’ voice signal obtained UBM, then get GMM model that waiting for recognition using MAP. Calculates each speaker’s log probability score, by analyzing and comparing distinguish different speakers, determining the speaker’s identity. Experiments show that the improved characteristic parameter TEO-MFCC compared to MFCC and MFCC+?MFCC, without increasing the computational complexity and improving the system recognition rate. In addition, in order to choose the most effective parameters for experiment, this paper also analyzes the dimension reduction and weighted algorithm of correlation distance Fisher criterion, Gaussian mixture model orders and the length of test voice on recognition results. Finally, the speaker recognition experiments has been done systematically with the two improved algorithms’ combination of the voice activity detection and feature extraction.