Study On The Difference Of Chinese Vowels Dysarthria In Parkinson’s Disease

The study of Parkinson’s dysarthria based on sustained vowels is one of the current research hotspots.However,due to the existence of differences in vowels,an effective diagnosis method for only one vowel is limited when patients with Parkinson’s disease cannot pronounce the vowel correctly.Therefore,this paper analyzes the differences of vowels through the classical acoustic features of Parkinson’s disease,eliminates the influence of vowel differences on Parkinson’s disease dysarthria from the aspects of features and samples,and proposes an effective diagnosis method for each vowel type of dysarthria.Firstly,the classical acoustic characteristics of Parkinson’s disease were used to analyze the different classification properties of Chinese vowels and the reasons for the differences.Classical acoustic features of Parkinson’s disease dysphonia were extracted from Chinese sustained vowel datasets,and the influence of vowel variances on diagnosis results and mixed diagnosis of multiple vowels were analyzed through experiments.Because the classical acoustic features of Parkinson’s disease only pay attention to the unilateral information in the time domain or frequency domain,the changes in the acoustic parameters of vowel pronunciation will affect the features,resulting in the heterogeneity of Chinese vowel performance.Secondly,aiming at the variances of Chinese vowels in classical acoustic features of Parkinson’s disease,a method to eliminate the differences of vowels based on fractional domain energy gradient features was proposed.Vowels are converted to fractional domain by fractional Fourier transform and energy gradient features are extracted from fractional domain energy spectrum.By using fractional Fourier transform,the time domain information and frequency domain information are combined in different directions and angles,and the optimal transformation order is selected to extract the energy change information from the direction with the greatest energy change rate,thus eliminating the difference of single time domain or frequency domain characteristics caused by vowel pronunciation.The experimental results show that the fractional domain energy gradient feature makes the classification accuracy of different vowels close,weakens the differences among Chinese vowel features,and obtains a set of features with strong applicability to multiple vowels,and improves the classification accuracy.Finally,in order to reduce the accuracy of multi-vowel mixed diagnosis due to sample differences,a two-stage vowel variances elimination method based on sample weighting and feature selection was proposed.The maximum information coefficient between samples was taken as the weight of sample weighted classification,and the maximum information coefficient and correlation coefficient were combined for feature selection according to the minimization principle.Sample weighting is used to reduce the contribution of samples with large variances in classification,and features with high correlation are selected to eliminate vowel differences by combining sample weighting and feature selection.The experimental results show that this method reduces the accuracy differences between the diagnosis of multi-vowel mixture and the diagnosis of unitary sound,and can further improve the classification accuracy of Chinese vowel,which proves the feasibility and advanced nature of this method.