基于原始波形的端到端阿尔茨海默症检测方法

Raw Waveform-Based End-to-End Alzheimer's Disease Detection Method

阿尔茨海默症(Alzheimer’s Disease,AD)是一种退行性疾病,随着病情加重,患者的语言能力逐渐减弱.目前已经有研究者使用梅尔谱图、梅尔频率倒谱系数(Mel Frequency Cepstral Coefficient,MFCC)等声学特征对阿尔茨海默症患者和健康人进行分类,但是对于使用神经网络从原始波形提取特征进行阿尔茨海默症检测还缺少进一步的探索.本文提出一种基于原始波形的端到端阿尔茨海默症检测方法.该方法使用一维卷积从原始波形中提取时间维度特征,并使用含有膨胀卷积的残差块提取更复杂的特征.为进一步提高性能,在残差块中引入挤压-激励模块.在全国人机语音通讯学术会议(National Conference on Man-Machine Speech Communication,NCMMSC)2021 AD数据集上,本文提出的模型在长音频测试集、短音频测试集分别达到了86.55%和81.35%的准确率,比基线系统分别提高了6.75%、7.35%.在INTERSPEECH2020 ADReSS数据集上,模型的准确率为66.67%,比基线系统提高4.17%.

Alzheimer's disease(AD)is a degenerative disease,as the disease worsens,the patient's language ability gradually decreases.Some researchers have already used acoustic features such as Mel spectrogram and Mel frequency cepstral coefficient(MFCC)to classify AD patients and healthy individuals,but there is a lack of further exploration on using neural networks to extract features from raw waveforms for AD detection.In this paper,we propose an end-to-end AD detection method based on raw waveforms.The method uses one-dimensional convolution to extract time-dimensional features from the original waveform and uses a residual block containing an inflated convolution to extract more complex features.To further improve performance,the squeeze-and-excitation block is introduced into the residual block.On the national conference on man-machine speech communication(NCMMSC)2021 AD dataset,the model proposed in this paper achieves 86.55%and 81.35%accuracy on the long audio test set and short audio test set,respectively,which is 6.75%and 7.35%better than the baseline system,respectively.On the INTERSPEECH2020 ADReSS dataset,the accuracy of the model is 66.67%,an improvement of 4.17%over the baseline system.