语音分离技术在阿尔兹海默症识别中的应用  

Recognition of Alzheimer’s disease by speech separation technology based on deep learning

作　　者：王学健 王杰[1] 王小亚 袁旻忞[3] 桑晋秋[4] 蔡娟娟[5] WANG Xuejian;WANG Jie;WANG Xiaoya;YUAN Minmin;SANG Jinqiu;CAI Juanjuan(School of Electronics and Communication Engineering,Guangzhou University,Guangzhou 510725,China;Guangzhou Women and Children′s Medical Center,Guangzhou 510168,China;Research Institute of Highway Ministry of Transport,Beijing 100088,China;University of Chinese Academy of Sciences,Beijing 100049,China;State Key Laboratory of Media Convergence and Communication,Communication University of China,Beijing 100024,China)

机构地区：[1]广州大学电子与通信工程学院,广州市510725 [2]广州市妇女儿童医疗中心,广州市510168 [3]交通运输部公路科学研究院,北京100088 [4]中国科学院声学研究所,北京100049 [5]中国传媒大学媒体融合与传播国家重点实验室,北京100024

出　　处：《中国传媒大学学报（自然科学版）》2022年第3期29-35,共7页Journal of Communication University of China：Science and Technology

基　　金：媒体融合与传播国家重点实验室(中国传媒大学)开放课题(SKLMCC2021KF014);国家自然科学基金(11974086,12074403);广州大学校内科研项目(YJ2021008)。

摘　　要：阿尔兹海默症的识别是预防与治疗该疾病的重要环节,目前的识别及进一步的诊断程序需要医疗专家进行全面检查,消耗大量的成本和时间。本文基于阿尔兹海默症早期认知障碍患者和确诊患者与正常人语言能力的差异,及语音分离模型的语言分类能力,在语音分离模型的基础上加入设计的语言障碍情况鉴别器,提出一种轻量化阿尔兹海默症深度学习识别方法,便于实现对这三种人群的识别,帮助医疗人员进行快速筛查。实验结果表明,本文使用的方法在MFCC特征集上的识别正确率可达84%,相比于基线系统提升约20%,且模型参数量仅有0.54M。此外,在频谱特征集合中,本文模型识别正确率提高约1.4%,参数量为0.23M。在梅尔频谱特征集合中,本文模型识别正确率也提升约4.4%,所需参数量仅为0.21M。The recognition of Alzheimer’s disease is an important step in the prevention and treatment of this disease.The current identification and further diagnostic procedures require thorough examinations by medical experts,which consume a great deal of cost and time.Based on the differences in language ability between patients with early cognitive impairment,diagnosed patients and normal people,as well as the language classification ability of the speech separation model,a lightweight deep learning recognition method for Alzheimer’s disease is proposed.By adding a designed language disorder discriminator on speech separation model,it is convenient to realize the recognition of these three groups and help medical personnel to conduct rapid screening.The experimental results show that the accuracy of the method can reach 84%in the MFCC feature set,which is 20%better than the baseline system,and the number of model parameters is 0.54M.In addition,for the Spectrum features set and Mel-Frequency Spectrum set,the accuracy of the model is improved by about 1.4%and 4.4%,and the parameters are0.23M and 0.21M respectively.

关 键 词：阿尔兹海默症 语音分离 鉴别器 深度学习 识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]