Research On Recognition Of The Pathological Voice Based On ANN And SVM

Along with social progress, economic development and the attendant improvement of life and the accelerating of the pace of life, the continuous enhancement of communication channels and capacity, the incidence of the common voice diseases is showing a rising trend. In the integrated emerging field of voice medicine, pathological voice recognition research based acoustic assessment and the use of signal processing means has gradually become a hotspot in many areas.Pathological voice recognition includes voice signal feature extraction and classification of voice samples. Based on this focus on the two most commonly used classification of Artificial Neural Network and Support Vector Machine, making full use of the Massachusetts Eye and Ear Infirmary open voice data, we have an offline research on pattern recognition and classification algorithm in view of 177 cases of common pathological voice mixed samples and 39 samples of normal voice. In this study, experiments are designed to analyze the feasibility of classification of the voice samples for ANN and SVM respectively. In the experiment design for ANN, the discussion and analysis of the corresponding data is in view of selecting the network structure parameters, such as the hidden layers, hidden nodes, and types of training function etc. In the experimental design for SVM, the discussion and analysis of the corresponding data is in view of choosing the type of nuclear function as well as optimizing the kernel parameters. In the end experimental design for the ANN and SVM comparison, the corresponding data analysis and discussion is in view of the degree of influence by the sample of both classifications and the stability of recognition rate.The results of this study indicate that both ANN and SVM classifier are applicable to the pathological voice samples pattern classification, with a higher rate of correct recognition. However, the ANN network model design and structural parameters selected often need more training and practical experience before been able to achieve optimal results; Compared to SVM, ANN is more sensitive to the training samples, and lack of good stability. Therefore, ANN classification is applied mainly to Off-line processing in the signal. On the other hand, within the optimized kernel function parameters interval by the grid enumeration, training model of SVM has good accuracy and stability of the final recognition rate. So the SVM classification can be applied to the need for processing in voice signals at rapid or real-time online occasions.