基于自我相关加工的抑郁症语音声学建模研究

For human, depression has long been recognized as one of the leading causes of disability and burden worldwide. In recent yesrs, research into the automatic and objective classification of depression, using various ways has gained popularity. Among these studies on depression, two main methods were involed: one is that the study of depressive based on biology, and the other one is that the study of depressive based on psychology. Although many scholars yielded substantial results from two aspects: theory and empirical analysis, clinical practice remains rooted almost exclusively on the opinion of individual clinicians. Current state-of-the-art diagnostic methods include interview style assessments. However, for clinical depression , these tests are subjective and single in nature , and lack an objective predictor of depressive. For this study, two basic problems must be solved: 1) how to build speech database based on the Self, in light of the psychological characteristics of self in depressed people; 2) How to construct acoustic model based on the traditional method of phonetics and experimental psychology method, and capture , analysis specific speech characteristics for performing depression prediction. Therefore, it is well known that the Self is not only the cognitive subject, but also the core of personality. We employ classial scientific psychology paradigms on abnormalities of self-related processing in patients from different dimensions of the Self, and adopt speech signal processing methods and Machine Learning methods for depressed speech. For this purpose, we propose a study on modeling for depressed speech from psychological perspective. The study results will provide a new technology and strategy for the diagnosis and early screening in depression.

抑郁症已成为损害人类身心健康的主要原因之一。目前，抑郁症的研究主要分为两类：分别是基于生物学和心理学的策略。尽管众多学者在抑郁症的理论研究和实证探索方面取得了丰硕成果，但是目前临床医生对其判别仍以主观量表为主，依赖主观判断，方法单一，缺乏有效的客观评价指标。对于本项目来说，我们需要解决两个基本问题：第一、如何针对抑郁症个体在心理自我方面的特点，实现建立基于自我的语音数据库方法；第二、如何结合实验心理学和语音学方法，提取分析潜在关联的声学特征参数，实现表征相应语音声学特征的模型。为此，本项目基于自我是认知的主体也是人格核心的特点，采用经典心理学范式，在自我的不同维度下观察抑郁症患者在自我相关加工方面的特点，针对抑郁症患者语音特征，运用语音信号处理和机器学习等方法，建立心理学自我相关加工和抑郁症语音表现力联合模型。研究成果将为抑郁症的客观测量和早期的筛查、诊断提供一种新的技术策略。

抑郁已成为损害人类身心健康的主要原因之一。虽现已有一些基于生物学和心理学的方法在理论研究和实证探索方面取得了丰硕成果，但纵观现今的客观识别表征方法还存在多项挑战性的问题与困难。如何解决这些问题，正是本项目的目标所在。区别于现有的研究方法，本项目以心理学自我相关加工视角分析与基于计算机机器学习表征建模为出发点，力求融合新的思路提出并研究新模型与方法，提供基于自我相关加工的抑郁语音声学建模方式的有效参考，形成一种尝试解决现有挑战性问题的可行且有效的方案，丰富自我与抑郁关系的研究策略。项目围绕研究计划，不仅按照设定的研究内容进行了项目实施，而且还在具体研究过程中扩展了相关研究。主要工作和成果如下：1、在抑郁语料的设计与建立方面。提出了一种基于心理学自我相关加工的抑郁语料库设计建立思路，在发声练习、语音朗读、自由响应和随意讲话等四个方面设计的基础上，又引入图片描述任务，并完善了基于特定问题的对话，采集了相应的抑郁语音语料数据。2、在基于自我相关加工的抑郁语音声学表征建模方面。首先提出了自我声音识别方法，通过自我声音识别考察抑郁个体自我识别能力的缺损，来探究抑郁个体的自我意识差异；其次，通过从元认知评价的视角来探索抑郁个体反思的自我加工，提出了基于自我反思的抑郁个体自我评价偏向表征方法；接着，提取语谱图输入卷积神经网络，构建了基于卷积神经网络的抑郁识别模型，进行二分类；并提出了基于残差卷积神经网络结合迁移学习算法进行抑郁识别的方法；最后，提出了一种运用卷积神经网络与双向长短时记忆网络进行抑郁识别的方法；实验结果表明，各方法识别性能指标达到要求。3、在精神健康延伸研究方面。发现在视听整合任务中，精神分裂症患者的我面孔会促进自我声音的识别等结论。项目组共发表或录用学术论文37篇，申请专利2项，获得软件著作权10项，与多位国内外学者开展了学术交流与合作。培养博士研究生10名，硕士研究生31名。