Research On Features Extraction And Recognition Algorithm For Pathological Voice

Voice is very important to the daily life and work of human beings.With the increase of people’s social activities and the change of living habits,the incidence of laryngeal diseases is becoming higher and higher.Therefore,the detection and treatment of voice has become the focus of relevant researchers.The essence of pathological voice detection and diagnosis is to convert the collected voice files into digital signals,extract their characteristic parameters,and compare the extracted feature parameters with the voice database parameters which are classified by the subjective judgment of the clinician,and then identify the categories of the tested samples by the recognition machine and thus achieve the disease.Among them,feature extraction and recognition machine is the key technology in voice recognition.How to select the features or feature groups and recognition machines which can fully express the voice characteristics and improve the classification accuracy of the pathological voice has always been the core issue of the relevant researchers.This paper mainly studies the optimization of feature and recognition machine of pathological voice,mainly including two aspects.1.On the basis of laboratory predecessors,we continue to explore the robustness of traditional acoustic and nonlinear features and their contribution to the pathological voice.Two different databases(Self-built clinical database and MEEI database)were tested in this paper.(1)In this paper,the method of feature ranking based on random forest was used to sort the contribution rate of 17 common features.The contribution of each feature to the recognition of pathological voice is judged by calculating the value of the decrease of average impurity and average accuracy rate of each feature.In order to further study the robustness and recognition contribution of selected features,an experiment of database crossover is carried out in this paper.The experimental results show that the characteristics of relatively high robustness,such as Hurst parameter,fuzzy entropy 2-Rényi entropy,attractor and fundamental frequency.Then random feature combination is carried out for these features with high robustness.In two different databases Support vector machine(SVM)and random forest are used for identification.Among them,the combined feature is:(1)Hurst parameter,attractor,fuzzy entropy,2-Rényi entropy,fundamental frequency;(2)Hurst parameter,attractor,fuzzy entropy,fundamental frequency;(3)Hurst parameter,attractor,2-Rényi entropy,fundamental frequency.These three groups show better robustness and recognition effect than other feature combinations.(2)In the experiment of feature combination recognition,(1)On the self-built clinical database,get the highest recognition rate is 97.33% and the average recognition rate is 89.12% based on SVM,using random forest to get the highest recognition rate is 98.00 % and the average recognition rate is 96.08 %.(2)On the MEEI database,using SVM get the highest recognition rate is 99.70 % and the average recognition rate is 97.77 %,the highest recognition rate was 100 %,the average recognition rate is 99.67 %.Combined with the experiments of single feature and combination feature,it is proved that the random forest is used as classifiers and show better classification accuracy and robustness than SVM.2.This paper also studies the features of different degrees of pathological voice in the self-built clinical database,and some combination features with good robustness were selected in the previous experiments.The ability of combination to distinguish normal voice,light voice,moderate voice and severe voice in multilevel database was analyzed.The highest recognition rates of four kinds of voice signals obtained by random forest were 98.72%,70.37%,79.07% and 87.88% respectively.The experimental results show that the experiment based on the characteristic combination of random forest can effectively detect different degrees of pathological voice,and also shows that the random forest has the characteristics of high classification accuracy.