Key Techniques Of Voice Diagnosis System

-IIIn recent years,method for diagnosing diseases by voice are widely concerned,because of the advantages of being simple,convenient,rapid,furthermore,there is no need to damage the body of the patient and the advantages of non-invasive examination.Although there are so many researches about sound diagnosis,there is still a lack of uniform sampling procedures and standards.Many studies,because of the lack of the number of sample sets,are inconclusive.Even for the different sampling parameters,the results of the study cannot be reproduced by other researchers.Research on this topic mainly focuses on voice diagnosis system acquisition and analysis technology,including the pronunciation content and duration,voice feature selection and dimension reduction,the key technical problems of acoustic diagnostic system of sampling frequency and quantization bits selection.To determine the standardized sample collection process,we need to select the appropriate sampling hardware equipment,including the microphone and sound card.The sampling process also includes the contents of pronunciation and duration of pronunciation,this paper recommended by the U.S.National vowel voice center,and taking account of the characteristics of pronunciation where we sampling,chooses 28 vowels.Combined with the actual situation,we choose 2 seconds as the duration of pronunciation.This topic divided diseases in database into the nervous system,lung diseases and vocal tract of three categories.In order to establish a sound diagnosis system of sampling frequency and quantization bits must be told.The classification experiment for different sampling frequency has been finished,according to the results,and thinking about the operation time and the storage space and other factors,this study pointed out the sampling frequency what’s recommended.The same analysis is carried out for different quantization bits,and the recommendation quantization bits is given.In order to select the appropriate features of sound and dimension reduction method,this paper makes a classification experiment on the commonly used voice features,and combines with the actual characteristics of the sound to make a choice.Among them,the classification accuracy of Mel’s coefficient is much higher than other features.In this paper,we choose 7 different dimension reduction methods,and choose the most appropriate dimension reduction method,which is the principal component analysis method with kernel function.All the diseases in the database have been carried out in the classification experiment,combined with the results and the pathology,analysis of various diseases separability.Among them,Parkinson’s classification accuracy rate is more than 87%,and the classification of cardiac arrhythmias,diabetes and lung cancer the accuracy rate has reached more than 80%.Finally,a prototype system is designed and implemented.The system includes ahigh fidelity sound signal acquisition module,a pathological feature extraction module,a disease classification module,and an analysis report output module.At present,this system can be used for the analysis of three kinds of diseases,such as nervous system,lung diseases and vocal tract.