Vocal Cord Pathological Speech Recognition With Non-traditional Feature Parameter

Due to vocal pathology caused by the surface irritation,tissue changes and other factors,the mobility,functionality,and shape of the vocal folds are affected resulting into irregular vibrations of vocal cords and poor overall voice quality.At present,the detection of vocal cords pathology is mainly through invasive laryngostroboscopy,which causes inconvenience to patients and depends on the subjective experience of clinicians,different doctors may judge the results differently.People hope to use computer tools to analyze voice signals in a noninvasive manner that help the objective diagnosis of vocal cords lesions and provide an objective standard for evaluation of voice quality.For these reasons,it is particularly important to use signal processing techniques and machine learning algorithms to calculate voice features automatically and recognize normal and pathological voices accurately.In order to evaluate the normal and pathological voices accurately,it is necessary to extract the feature parameters that can well reflect the pathological characteristics.Therefore,this thesis began with the analysis of the linear and nonlinear characteristics of voice signals,and on the basis of traditional feature parameters,a series of transformations of voice signals were made to obtain nontraditional feature parameters.Different feature parameters were input into support vector machine for modeling and recognition,and the recognition ability of each parameter to vocal cords pathological voices was judged by the classification accuracy.In this thesis,the following conclusions were obtained through theoretical analysis and simulation experiments:1.Compared with the traditional linear acoustic parameters,the extraction of nontraditional linear feature parameters in this paper did not need to accurately judge the pitch period,and it's easy to implement and apply.The results of support vector machine recognition showed that the nontraditional linear feature parameters were better than traditional linear parameters in identifying normal and pathological voices.These nontraditional features improved the ability of traditional features to analyze pathological voices to a certain extent.2.Four-layer wavelet packet decomposition was used to extract the nonlinear feature parameters in different frequency bands,and the nontraditional nonlinear hierarchical features were obtained.By comparing the recognition rate of feature parameters before and after layered,the hierarchical feature parameters improved the ability of recognizing pathological voices,and the dimension-reduced optimized feature parameters contained more useful information,which was more conducive to the identification of pathological voices.3.Different parameters had different recognition ability to normal and pathological voices.By comparing the average recognition rate of all parameters under different classifiers,it was found that the highest recognition rate was 98.20%of hierarchical optimized fuzzy entropy in nonlinear feature parameters,followed by 98.19%of the spectral flatness of the residue signal in linear parameters.Considering that the dimension of hierarchical optimized fuzzy entropy was higher than the spectral flatness of residue signal,and the time cost was higher,it was better to choose the spectral flatness of residue signal when distinguishing normal voices from pathological voices.4.By comparing the recognition rate of paralysis and polyp voices with different parameters,it showed that the nonlinear hierarchical optimized feature 16-dimension Hurst exponent had the best recognition rate,which was 96.97%,it can distinguish paralysis voices from polyp voices accurately and effectively.The results provided a possibility for noninvasive detection of different types of pathological voices.5.The nonlinear feature parameters had more advantages in distinguishing different types of pathological voices,which was related to the pronunciation mechanism of vocal cords pathological voices.The abnormal vibration of vocal cords caused by laryngeal lesions leaded to the increase of turbulence of voices,and the nonlinear behaviors were more obvious.Therefore,nonlinear analysis method was more suitable to study the characteristics of different types of pathological voices.