Robust Speaker Recognition Based On Sparse Coding

In recent ten years,compared with the traditional identity authentication methods,biometric identity recognition technology has attracted the extensive attention and in-depth study.Most notable,speaker recognition technology is a new research direction,it can recognize the identity of speakers through their voice,it is the same as the authentication technologies in the field of biometric technology such as current widely used fingerprint recognition,face recognition and iris recognition.Compared with other biometric authentication technologies,speaker recognition has great advantages,such as high popularity of biometric devices,and no invasion in the whole authentication process.Therefore,the application of speaker recognition to reality is significance.However,the lack of voice and environmental noise in practical applications put forward high requirements on the robustness of speaker recognition method.At present,almost all the speaker recognition methods are based on the idea of model matching,so the research focus on the speaker model.In the common models,GMM(Gauss Mixture Model)is very popular.Especially,the models develop from the GMM have made a very good recognition effect.Such as GMM-UBM(Gaussian Mixture Model-Universal Background Model)and GMM-SVM(Gaussian Mixture Model-Support Vector Machine).But in the presence of noise and small amount of speech data,the recognition rate of these speaker recognition models decrease significantly.Although there are many improved methods have been proposed,but the robustness of speaker recognition can't meet the actual requirement very well.To solve the above problems,this paper mainly studies robust speaker recognition methods under the condition of small amount of speech data and environment with noisy train data or test data.Firstly,for the problem of only a small amount of speech data available,a speaker recognition method based on sparse coding is proposed.In the trainingphase,a dictionary is trained for each speaker,and then the score is recognized according to the reconstruction error.In the absence of noise and insufficient data,the classical GMM-UBM and GMM-SVM methods are compared with the proposed method.Then,on the basis of speaker recognition method based on MCA(Morphological Component Analysis),a new speaker recognition method based on GMM-UBM is proposed.The new method training a background dictionary with all of training speech data,the dictionary of each target speaker can be got from optimization of the background dictionary.Adding a concept of noise dictionary and these three dictionaries finally spliced into a large dictionary for sparse decomposition,so that the speaker recognition could have immunity to noisy test data.With the help of S-SGK(Sparse Sequential Generalization of K-means)which can train a dictionary with noisy signal,a robust speaker recognition method for the condition of noisy train data is proposed.This paper makes a lot of experiments,and the results show that the speaker recognition method based on sparse coding have a better performance than GMM-UBM and GMM-SVM recognition method when the voice data is not sufficient.Under the condition of synthetic noisy test speech data,speaker recognition method based on MCA is proposed has higher recognition accuracy than some other common methods.With the synthetic noisy trian speech data,the method based on S-SGK also has better performance than some common method.