The Algorithm Of Speaker Identification In Noisy Environment Base On GMM/SVM

Language is the most important tool for human intercommunion. Speech signal as the carrier of language embodies much information in different level. The information about speaker can be used for identify the people who is speaking or whether he is the specified people. In modern day, automatic speaker recognition has performed quite perfect in the low noise and low distortion. But the mismatch between the training data and the test data that result from all kinds of noise in real environment make the speaker recognition rate dramatically declined. So improving the performance in the noise environment is the key for the system come to practice from laboratory.The technology of speaker recognition is composed of feature extraction and pattern classification. This paper studied feature extraction and robustness by analyzing pronunciation organ and hearing organ. In addition, some primary classifiers are intensively researched. All of works focus on text-independent open-set speaker recognition in noisy environment.Considering information entropy that is comprehensively applied to code theory represent average unconfirmed information source, the entropy of speech and the entropy of noise must be different. This paper adopts speech endpoint detection method based on entropy function. The experiment shows the spectrum entropy performed much well in low SNR and unconfirmed noisy condition. Further, a method based on dynamic threshold is proposed to detect speech endpoint.Considering noise frequency band rarely covers the whole scope of speech, this paper adopts multiple subbands feature extraction and uses sub-cepstrum based Teager energy in every subband. Furthermore, a hybrid system of Support Vector Machine (SVM) and Gaussian Mixed Model (GMM) optimized by AdaBoost algorithm is introduced. Firstly, this system filters out the speaker not included in the training set by using SVM. Then, weights the decision results of the training set to highlight subband features that affected the recognition results much more than others. By this way, the noise impact on the outcome of the identification is reduced. Finally, the optimized GMM is used for recognition. The experiment result shows this system performed still well in lower SNR condition.