Statistical Analysis of Local Gradient in Mel Transform Domain for Parkinson’s Dysphonia

Dysphonia analysis of Parkinson’s disease is the basis of information analysis for early diagnosis of Parkinson’s disease based on speech. In recent years, with the deepening of research, Mel transform domain information shows more and more advantages in this field. At the same time, the improvement of classification performance by extracting structural features is increasingly apparent. This paper proposes a method for local gradient statistical feature extraction in Mel transform domain from the point of the structure of Mel transform domain information of speech signals of people with Parkinson’s disease. Firstly, the speech signal is converted into the energy signal in the time-frequency transform domain by the method of Mel frequency transformation, and the energy spectrum is represented visually. Then, the energy data are processed by sliding window, and the local structure information of the Mel transform domain is obtained by calculating the gradient and angle of each energy point in the detection window. Finally, the gradients of the energy points of all detection windows are calculated according to the angles to obtain the local gradient statistical features, which represent the change of energy value in Mel transform domain. The results of the experiments performed on different datasets by different classifiers show that compared with the methods of Mel transform domain analysis, cepstrum analysis and deep learning, the local gradient statistical features in Mel transform domain are superior to them in classification accuracy and sensitivity, thereby verifying the effectiveness of the local gradient statistical feature in the dysphonia analysis of Parkinson’s disease.