Vocal Cords Diseases Detection By Multi-band Nonlinear Analysis And Perception Polyspectra Entropy

Voice is the most natural mean of communication in contemporary society, and it willshow its superiority and necessity in the field of intelligent human-computer interactionwith the development of science and technology. Voice diseases rise due to environmentalproblems, social and occupational stress. As the most important vocal organ, vocal cordsdiseases are main cause of voice problems. Pathological voices diagnosis by acousticanalysis has become a research focus, because it is noninvasive, fast, objective andsupporting remote detection etc.This paper mainly analyses the vocal cords diseases voice. Limitation and deficiencywill arise when traditional features describe influences of vocal cords diseases to voicessignal, multiband nonlinear analysis and perception polyspectra entropy are proposed to improvethe vocal cords diseases recognition rate. Specific studies are shown as follows:(1) The influences of vocal cords diseases to voice signal are introduced andreflections and limitations of traditional characteristics are also mentioned, and thencorrelation and difference analysis is carried out for feature selection. Gammatone filterbanks are introduced in detail, whose frequency response can simulate human earcharacteristics well. Multi-band largest Lyapunov exponent is proposed when it iscombined with nonlinear dynamic characteristic largest Lyapunov exponent, which candescribe nonlinear characteristic of voice signals in each frequency band precisely.Polyspectra calculation for each Gammatone band together with energy entropy analysisare implemented to propose perception polyspectra entropy characteristics, which is a goodrepresentation of non-Gaussian nature of voice for each frequency band. Recognitionexperiments of normal voices with vocal cord diseases voices and vocal cord diseases withhyperthyroidism thyroid disease voices show that the proposed features achieve higher accuracy than traditional features.(2) Kernel principal component analysis is utilized to reduce redundancy betweendimensions, and the form of kernel function and kernel parameters influence theoptimization effects directly. The paper puts forward a kernel parameters selectionalgorithm based on Gaussian maximum likelihood rule. Recognition experiments show thebest performance of proposed kernel function.(3) The choicely traditional features and proposed features are merged as a newcharacteristic vector to comprehensively describle influences of vocal cords diseases tovoice signal, which economize complementary characteristics of different features. Theaverage recognition rate reaches97.83%when fused feature used for normal voices andvocal cord diseases voices recognition system, and average recognition rate of80.81%forvocal cords diseases voices and hyperthyroidism thyroid disease voices recognitionsystem.(4) Deficiencies are pointed out and the correspondent improvements and suggestionsare proposed finally.