Recognition And Study Of Pathological Voice Based On Nonlinear Dynamics Using Gaussian Mixture Model/Support Vector Machine

The pathological vocal cords and voice should be diagnosed at an early stage, it is well known, and the lesions of vocal cord can cause the changes in the speech signal. Therefore, the sound signal can be used as an important objective non-invasive tool to diagnose these lesions. In recent years, non-destructive diagnosis of voice pathology is an important issue, the classification of normal and pathological voice, as an auxiliary treatment, has been more and more researchers’ attention. Currently Clinicians to use existing electronic equipment, such as EMG, tongue to move the current tracings and the electronic endoscope to detect and diagnose the vocal cord lesions. However, each of the test methods are invasive, require professional doctor to analyze the acoustic parameters of voice signal, and ultimately the results can not escape subjective judgment. As computer science and technology development, it advances pathological voice objective, non-invasive, painless diagnostic. However, the objective diagnosis of pathological voice:and intelligent recognition, can not be well applied to the clinical detection, but also there is a certain difference from the clinical detection.And these electronic instruments extract the general traditional acoustic characteristics parameters of speech signal, such as fundamental frequency, formant, Mel frequency Cepstral coefficients, etc. These traditional acoustic parameters are extracted by using traditional linear analysis techniques. It has some limitations for on-periodic and chaotic voice signal. And the aerodynamic and acoustic theory suggests that the voice signal exists a chaotic mechanism. it is a complex nonlinear process. The traditional acoustic analysis methods ignore the nonlinear characteristics of the voice. It seriously impacts on the effectiveness of the characteristics. Fractal theory and Chaos Theory of Nonlinear Dynamics technology can effectively describe the irregular and non-periodic signal.In this study, it quantities the nonlinear characteristics of the voice using the analytical methods of nonlinear dynamics, extracts the nine-dimensional non-linear characteristic parameters of the voice, to make up for the deficiencies of the traditional analysis methods. And neural network feature selection is used to sift the nonlinear characteristics, obtain7dimensional feature parameters:Hurst parameters,the second order Renyi entropy, correlation dimension, Kolmogorov entropy, largest Lyapunov exponent, box-counting dimension and intercept.Using of7-dimensional characteristic parameters and different from the traditional pattern recognition methods-Gaussian mixture model and support vector machine, normal and pathological voice is recognized automatically and non-invasively The database of151cases of voice in39normal and36cases of pathological voice as the test set, a better recognition rate of GMM and SVM is obtained respectively:98.61%and97.33%. The experimental results show that the analytical methods of nonlinear dynamics can effectively analyze normal and pathological voice, using GMM and SVM recognition methods, and can achieve a higher recognition rate.Also compared with the traditional acoustic characteristic parameters, fundamental frequency, formant and MFCC, the results show that compared that with the traditional acoustic characteristic parameters, the parameters of the nonlinear dynamical characteristics are more effective. Using the combination of three-dimensional nonlinear characteristic parameters and fundamental frequency, formant that they can achieve a better recognition rate, the recognition rate achieved is better than traditional acoustic parameters, which indicates that the nonlinear dynamics parameters can compensate for the deficiencies of the traditional acoustic parameters and it can improve our ability of understanding and analysis of pathological vocal voice.