基于深度学习的音频抑郁症识别  

AUDIO DEPRESSION RECOGNITION BASED ON DEEP LEARNING

作　　者：李金鸣 付小雁[1,2] Li Jinming;Fu Xiaoyan(College of Information Engineering, Capital Normal University, Beijing 100048, China;Beijing Key Laboratory of Electronic System Reliability Technology, Beijing 100048, China)

机构地区：[1]首都师范大学信息工程学院,北京100048 [2]电子系统可靠性技术北京市重点实验室,北京100048

出　　处：《计算机应用与软件》2019年第9期161-167,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61876112,61601311,61603022)

摘　　要：抑郁症以显著而持久的心境低落为主要临床特征,是心境障碍的主要类型,严重影响人们的日常生活和工作。研究人员发现,抑郁症患者与正常人在言语方面存在明显差别。提出一种基于卷积神经网络和长短时期记忆网络的音频抑郁回归模型(DR AudioNet)。从特征设计和网络架构两方面进行研究,提出多尺度的音频差分归一化(MADN)特征提取算法。MADN特征描述了非个性化讲话的特性,并根据音频段前后相邻两段的MADN特征设计基于DR AudioNet优化的两个网络模型。实验结果表明,该方法能够有效地识别抑郁程度。Depression is characterized by significant and persistent low mood, and is the main type of mood disorder, which seriously affects people s daily work and life. The researchers found that there were many significant differences in speech between normal people and people with depression. This paper presented an audio depression regression prediction model(DR AudioNet) based on convolutional neural networks and long short-term memory. Through research on feature extraction and network architecture, a multiscale audio delta normalization(MADN) feature extraction algorithm was proposed. MADN features described the characteristics of non-personalization speech, and two network models based on DR AudioNet optimization were designed according to the MADN feature of the two adjacent segments. The experimental results show that the method can effectively identify the degree of depression.

关 键 词：抑郁症自动诊断 语音信号处理 深度学习 音频特征 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]