Research On Strategies Against Abnormal Speech In Voiceprint Recognition System

Due to rapid development in technology of bio-information, using password or token toaddress various types of access control system or security system traditionally, which isdifficult to satisfy users needs for more secure and more convenient. Voice biometric featurewas widely used in the field of speaker recognition.When under quite or normal speech, theycan get higher recognition rate. However, when face various types of abnormal voice such ascomplexity background noise, change the manner of speaking or customary voice deliberately,vocal organs lesions voice, etc, the performance of all type of speaker recognition systemdecline sharply. So abnormal speech as a research object of speech signal processingtechnology has been concened widely, and developed into one of hot topic.In real life, speaker's voice can be produced pathological abnormal phenomena veryeasily, such as: cold, anti-acid, laryngitis, sore throat, stuffy nose, drunk, etc. In order toimprove the universal and to reduce the rejection rate of abnormal voice, the pathologicalabnormal voice such as cold voice is our object to study, the thesis focuses to explore a rapidpathological abnormal continuous speech detecting methods, and abnormal feature weightingdynamic for text-independent pathological abnormal continuous speech, and optimal speakercommon vector extraction. The main contribution of this thesis is as follows:1) A pathological abnormal continuous speech detecting algorithm based on correlationdimension is proposed to solve the problems, that artificial set optimal sampling delay andoptimal embedded correlation dimension algorithm can not descript the complexity of diseasecontinuous speech system objectively. A reasonable sampling delay range which can be obtainby analysis frequency of the signal to solve the pre-set the optimal sampling delay, and then,in the reasonable sampling delay rang, by adjusting the sampling delay to obtain itscorresponding embedded correlation dimension, and finally by equate error rate analyzing forembedded correlation dimension to obtain optimal chaotic parameter. The experimental showsthat the distinguish rate between normal and abnormal speech of the proposed is75.6%,compared with the GMM-SVM, Shimmer, pre-set sampling delay and fixed embeddeddimension algorithm, SHR and Jitter the distinguish rate increased by7.8%,9.3%,18%and20.4%, respectively.2) To solve drawbacks that the commonly-used weighting algorithm is inefficient intracking the abnormal feature of abnormal speech dynamically, an abnormal feature weightingalgorithm for abnormal speech is proposed. In this algorithm, the K-L distance and theEuclidean distance are used to measure the differences between a given test speech and the normal speech feature templates. Two weighting factors are used to weight the MFCCfeatures of the test speech, and the weighted MFCC features are input in the Gaussianmixture model for abnormal voice speaker recognition. Experimental results show that theglobal recognition rates of the speaker recognition algorithms based on the K-L weighting andthe Euclidean weighting are46.61%and42.25%, respectively. Compared with the traditionalweighting and without weighting algorithm, the speaker recognition ratio were increased by6.93%,10.25%and2.57%,5.89%. In slightly abnormal voice speaker recognition, thespeaker recognition rate of the K-L-weighting algorithm was83.77%, compared withtraditional weighting algorithm and without weighted algorithm the speaker recognition ratewere increased by16.51%and17.53%3) From exploring the common characteristics of normal voice and disease voice, anabnormal voice speaker recognition algorithm based on common vector is proposed toovercome objectively defects to set common vector parameters. Adaptive common vectorparameters search was used in the training period which makes the speaker recognition rateget maximum values. In the test period, the optimal common vector parameters used toextract a common vector for testing voice, and then SVM classifier used to speakerrecognition. The experiment results show that the speaker recognition rate of the proposed forslightly abnormal voice is85.4%, compared with the GMM, the SVM and combine commonvector with GMM, the speaker recognition were increased by16.9%,15.2%and3.2%,respectively. In slightly abnormal voice speaker recognition, the speaker recognition of thisproposed is51.8%, compared with the GMM, the SVM and combine common vector withGMM, the speaker recognition were increased by10%,8.6%and2%.4) To solve the problems that the performance of common used clustering algorithm isheavily dependent on pre-set optimal clustering parameters which are often difficult to obtain.Try to use soft-decision likelihood to speaker clustering without pre-set clustering parameters,a speaker clustering algorithm based on minimal GMM tracking dynamically is proposed.First, suspected clustering speech set was built by tracing the minimal Bhattacharyya distancebetween two Gaussian Mixture Models, and then likelihood was used to speaker verificationfor each two speech sets in suspected clustering speech set. When two speech sets wereconformed from the same speaker, then used the likelihood to verify each utterance from onespeech set is truly speak by the other speech set's speaker again. Experimental results showsthat theFs coreof this proposed is69.08%, compared with pre-set a better clusteringparameters of the K-means and ISODATA, theFs corewere increased by1.99%and0.95%, compared with one layer likelihood verification, the average clustering purity of two layerwas increased by5%,the averager speaker purity was increased by8.8%and theFs corewasincreased by7.08%.