Recognition And Study Of Pathological Voice Based On Vowel/a/and/i/

Voice is closely related to personal quality of life and happiness.How to detect pathological voice of the patients with voice lesions more effectively,determine lesion degree and evaluate the effect of the treatment are very important.The essence of detection and diagnosis of pathological voice is to convert voice samples into digital signals,extract the feature parameters,compare extracted features and the ones in voice database subjective judgment by clinicians,and at last determine the category of test samples.In the last decades,some studies have provided new methods,acoustic analysis,which is used for evaluation,postoperative rehabilitation and voice diseases diagnosis.Automatic detection and recognition of pathological voice will be realized.The essential features of pathological voice are characterized by feature parameters.Short-time analysis technology is adopted to extract the traditional feature parameters because of the short-time stationarity of voice signal.Traditional features:fundamental frequency,Mel cepstral coefficients(MFCC),linear prediction cepstrum coefficient(LPCC),frequency perturbation,amplitude perturbation,formant etc,are used widely to pathological voice recognition system.However,aerodynamic and acoustic theory indicates that chaotic mechanism exists in the voice signal,which is a complex nonlinear process.Traditional acoustic analysis methods ignore the nonlinear characteristics,while nonlinear dynamics technology can describe the irregular and non-periodic signal effectively.In this study,nonlinear characteristics including attractor,Poincare section mapping,sample entropy,fuzzy entropy,multi-scale entropy,Renyi entropy,Shannon entropy,intercepts,box-counting dimension and Hurst parameter were extracted to analyze the pathological features of pathological voice qualitatively,quantitatively,as feature vectors of identifying pathological voice based on SVM.A comparative study of pathological voice based on traditional acoustic characteristics and nonlinear features was carried out,and the results show that nonlinear characteristic can well distinguish between healthy and pathological voices,especially fuzzy entropy,best rates of 94.67%for/a/,87.58%for/i/.Recognition rate for/a/were all higher than that for/i/but multi-scale entropy,which indicates pathological voice of/a/can be recognized more easily and/i/can be affected by vocal cords compensatory ability more.In addition,this study proposed a method to make the parameter contain more pathological information by combining parameters as some criterion.A classifier based on SVM was implemented to evaluate the parameter.Global rates of 97.33%for/a and 92.41%for/i/are obtained,which proves the effectiveness of the combination features.At the same time,this paper studied the pathological character on different levels of pathological voice quality.Nonlinear dynamics technology was used to extract nonlinear features.Comparing the separation ability that nonlinear single feature or combination features have on healthy,light pathological,moderate pathological,severe pathological voice of vowel/a/and/i/was performed based on SVM and GMM.Best classification rates for these four kinds of voice of/a/were:96.10%,82.89%,85.33%,and 100%using single nonlinear feature based on SVM,the same ones of/i/being 85.71%,85.33%,77.63%,and 95.56%.Global success rates of 96.67%,78.89%,86.67%,and 100%for/a/,97.78%,77.78%,75.56%and 96.67%for/i/were obtained by using combination features based on SVM.When GMM was as the classifier,best rates were 91.03%,81.48%,65.12%and 100%for/a/,96.15%,77.78%,93.02%and 90.91%for/i/.Results obtained verify nonlinear dynamic characteristics can detect different levels of pathological voice effectively,and provide reliable theoretical basis for clinical diagnosis of laryngeal diseases.