基于深度学习的帕金森患者声纹识别

Voiceprint recognition of Parkinson patients based on deep learning

90%的帕金森病(PD)患者存在声带紊乱,提出一种利用WMFCC提取患者的声纹特征、DNN识别并分类的方法用于区分PD患者和健康人。通过计算患者声纹中倒谱系数的加权和系数,解决高阶倒谱系数小、特征分量对音频的表征能力差等问题。DNN训练并分类识别有效地提高系统精度,使用MBGD优化算法降低损失函数的计算量,提高系统训练速度。针对PD database中样本测试分类,结果相较传统SVM等方法,该方法提高了判别PD准确率,可达87.5%,为PD患者早期快速辅助诊断提供了良好的解决方案。

90% of the Parkinson disease (PD) patients suffer from vocal cord disorders. Therefore,a WMFCC method for extracting the voiceprint features of patients,DNN recognition and classification was proposed to distinguish PD patients from healthy people. By calculating the weighted sum coefficient of cepstrum coefficients in the patient ’s voiceprint,the problems that the high-order cepstrum coefficients are small and the feature component ’s representation ability to the audio is weak were solved. Through the data training,classifying and recognizing procedures using the DNN technology,and by using the MBGD optimization algorithm,the amount of computation of the loss function was reduced,thus increasing the training speed of the system. By testing and classifying the samples in the PD database,it shows that this approach improves the accuracy of the diagnosis of the Parkinson disease compared with the traditional SVM method and other methods. The accuracy rate can reach 87.5%. This approach provides a solid solution for early quick auxiliary diagnosis of Parkinson disease.