Recognition Of Pathological Voices And The Design Of Voice Recognition System Using Feature Parameters Optimized By Correlation Fusion

Language is an important way of communication.Clear and loud voice is the basis of smooth communication.In daily life,it is unavoidable to produce some unhealthy vocal behaviors,such as shouting,excessive use of voice,etc.These behaviors affect the voice health to varying degrees,while minor ones cause abnormal tones and quality of voice,and heavy ones cause vocal cord disorders and organic lesions,resulting in voice diseases.Usually,it is difficult for patients to pay attention to the early mild voice disease.Only when the diseases are very serious,patients will choose to see a doctor.At present,the main way to diagnose voice diseases is to go deep into patients' throats and observe the specific situation with visual equipment,which can easily bring physical discomfort and psychological pressure to patients.If there is a way to find abnormal voice when symptoms are mild,the unhealthy pronunciation behavior can be corrected in time,the voice health can be easily and quickly restored;or if there is a way to accurately judge the degree of voice abnormality before the vocal cord has serious lesions,and carry out the next step of diagnosis and treatment as needed,it may be able to avoid the aggravation of the patient's condition.Therefore,how to monitor voice abnormality conveniently in the early stage,and how to diagnose voice disease by noninvasive and digital means have become a research issue with strong scientific and practical significance.To solve this problem,a strategy of correlation analysis is proposed,which combines the optimized nonlinear feature to identify polyps and paralysis,and a noninvasive voice health monitoring and disease diagnosis system is designed based on simulation experiments.The main work and results of this paper are as follows:1.By reading literatures on voice recognition,it is found that using nonlinear features to identify pathological voice is more scientific and reasonable,because abnormal vocal cords produce nonlinear phenomena that can not be characterized by linear feature in the transient process of vocalization.The ability of identifying pathological voice by a single feature parameter is limited,it is found that feature fusion,feature remapping can fully show the characteristics of pathological voice signal and get good recognition accuracy.Therefore,six commonly used nonlinear features are selected in this paper:sample entropy(SE),fuzzy entropy(FE),permutation entropy(PE),Hurst index,Teager energy operator,L-Z complexity(LZC).The pathological voice signal is decomposed by a wavelet packet,and six features are extracted hierarchically for principal component analysis(PCA).According to correlation analysis,optimized features are fused to identify polyps and paralysis.2.Using voice linear prediction residual signal spectrum flatness(SF)to classify healthy and pathological voices,100%and 97.22%classification accuracy were obtained on support vector machine(SVM),respectively.The health and pathology voice classification model is then applied to recognize voice samples from the third speech database,the overall accuracy is more than 83%,and the recognition of healthy and pathological voices is preliminarily achieved.3.Two kinds of different pathological voice recognition simulation experiments are compared:six nonlinear features extracted from voice signal by traditional pathological voice recognition are simulated separately,and it is found that different features have great differences in classifying different pathologies.After fusing the optimized features,the pathological recognition rate is significantly improved.The overall accuracy of Hurst+LZC group and FE+LZC group was over 97%.This shows that the fusion features can highlight the effective details of pathological voice signals,and proves the effectiveness of the fusion strategy.4.Based on the results of the theoretical research in this paper,we designed a noninvasive voice health monitoring and disease diagnosis system.The main window complete the functions of user information input,voice real-time recording,voice health monitoring/diagnosis,information and result preservation.The dialog window displays the user information set and it can be called out by clicking the "medical record query"button in the main window.It can complete the functions of querying historical medical records and returning medical records to the main window."Health monitoring" function is designed by using the health and pathological voice recognition model in this paper,and the "pathological diagnosis" function is designed using the training model of the combination of optimized FE and LZC in this paper,which forms a complete closed loop from theoretical research to practical application.