Research On Feature Extraction For Pathological Voice

Laryngeal diseases can have different degrees of impact on the patient voice,so voice recognition technology can be used to assist in diagnosis and treatment of such diseases.Feature extraction is a key pathological voice recognition technology,the effectiveness of the extracted features directly affects the pattern recognition classification.Due to the complexity of the sound system,voice recognition,which is currently based on pathological parameters of acoustic characteristics,still does not meet the requirements of clinical recognition in spite of some progress.Therefore,a more efficient extraction of pathological voice features to improve the classification accuracy of pathological voice,has become particularly important.This paper focuses on pathological voice feature extraction technology carried out relevant research,including the following two aspects:1.An exploration of the contribution of traditional features and nonlinear features to pathological voice recognition is addressed in this paper.A total of 20 kinds of features that have previously been shown to be effective for pathological voice recognition problem are used in this experiment.A saliency measure by the BP neural network is then used to evaluate the contribution of these 20 kinds of features to the recognition problem.Then,the best set of features for pathological voice recognition is selected to perform the identify problem.With the support vector machine(SVM),we get the highest recognition rate of 98.67%,with the average recognition rate of 88.66%.2.In order to further improve the accuracy of normal and pathological voice classification,this paper proposes the use of wavelet technology voice signal preprocessing before feature extraction method is in principle based on the mechanism of voice.By analyzing the mechanism of vocal,it is known that voice disease mainly occurs in the vicinity of the vocal cords,so the process is a response by the removal of the voice channel to achieve improvement in recognition rate.Wherein by using a binary discrete wavelet transform to achieve improved fuzzy entropy,frequency perturbation,2nd Renyi entropy feature recognition rate,especially second order Renyi entropy,the highest recognition rate of 89.33%up to 97.33%,the average recognition rate from 79.19%up to 91.32%.Since the binary discrete wavelet transform high-frequency decomposition is not fine for this paper also proposes an alternative method using wavelet packet transform binary discrete wavelet transform.By using wavelet packet transform technology to improve the recognition rate for various non-linear characteristics,and each combination of features to improve the recognition rate,and finally,the use of Hurst parameter and box dimension as the feature vector sets in order to achieve best recognition rate.To be exact,the highest recognition rate is 98.67%and the average recognition rate is 93.92%.