注意力残差模型的语音抑郁倾向识别方法  

Speech Depression Tendency Recognition Based on Attentional Residual Network

作　　者：鲁小勇 石代敏 刘阳 原静仪 董强利[4] 马秀云[4] LU Xiao-yong;SHI Dai-min;LIU Yang;YUAN Jing-yi;DONG Qiang-li;MA Xiu-yun(School of Physics and Electronic Engineering Northwest Normal University,Lanzhou 730070,China;Internet Education Data Learning Analysis Technology National and Local Joint Engineering Laboratory,Lanzhou 730070,China;Key Laboratory of Behavior and Mental Health of Gansu Province,Lanzhou 730070,China;Department of Mental Health,Second Hospital of Lanzhou University,Lanzhou 730030,China)

机构地区：[1]西北师范大学物理与电子工程学院,兰州730070 [2]互联网教育数据学习分析技术国家地方联合工程实验室,兰州730070 [3]甘肃省行为与心理健康重点实验室,兰州730070 [4]兰州大学第二医院心理卫生科,兰州730030

出　　处：《小型微型计算机系统》2022年第8期1602-1608,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(31860285,31660281)资助.

摘　　要：采用语音信号进行抑郁倾向识别具有重要的现实意义.针对语音抑郁倾向识别使用深度神经网络方法结构复杂和传统机器学习方法需要手动提取特征及识别率低的问题.本文提出了一种结合残差思想和注意力机制的模型,首先基于心理学自我参照效应(Self-reference Effect,SRE)实验范式设计了抑郁语料,进行语音数据集标注;然后将注意力模块引入残差单元中,利用通道注意力学习其通道维度上的特征,空间注意力反馈其空间维度的特征,并将两者结合得到注意力残差单元;最后堆叠单元构建基于注意力残差网络的语音抑郁倾向识别模型.实验结果表明,与传统机器学习方法相比,该模型在抑郁倾向识别上获得了更优的结果,可满足抑郁倾向识别应用的需求.It is of great practical significance to use speech signals to identify depression tendency.In view of the complex structure of deep neural network method and the traditional machine learning method need to manually extract features and low recognition rate in speech depression recognition.This paper proposed a model combining the residual idea and the attention mechanism.Firstly,a depression corpus is designed based on the Self-reference Effect(Self-reference Effect,SRE)experimental paradigm of psychology to annotate the speech dataset.Then the attention module is introduced into the residual units,and the channel attention is used to learn the characteristics of the channel dimension,and the spatial attention is used to feedback the characteristics of the spatial dimension,and the attention residuals are obtained by combining the two.Finally,the stack unit is used to construct the recognition model of speech depression tendency based on attention residuals network.Experimental results show that,compared with the traditional machine learning methods,the proposed model achieves better results in the recognition of depression tendency,and can meet the needs of depression tendency recognition.

关 键 词：抑郁倾向识别 残差神经网络 注意力机制 Mel频率倒谱系数(MFCC) 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]