基于修正协作表示的语音重度抑郁症检测

发布时间：2018-04-22 11:15

本文选题：重度抑郁症 + 稀疏表示　；参考：《吉林大学》2017年硕士论文

【摘要】：重度抑郁症(MDD)是一种常见的精神紊乱疾病。患者呈现长期过度伤心、消极情绪及认知障碍,甚至自杀,严重影响其身心健康。第二版贝克抑郁量表中,MDD评分值范围是29~63。MDD患者语音通常具有单调、低沉、无生命力的迹象。同健康者语音相比,抑郁症语音蕴含的音源、系统和韵律信息具有本质差别。现有研究表明通过捕捉声学特性变化完全能够有效检测MDD。然而,抑郁症语料库的有限性和分布不平衡等已经成为重度抑郁症检测的主要障碍。传统分类模型无法满足需要,如支持向量机(SVM)、高斯混合模型(GMM)和稀疏表示分类器(SRC)等。当各类训练样本充足时,GMM性能良好。SVM虽然适合小样本分类问题,但要求各类样本数量应该平衡。当样本足够多时,稀疏表示以1l范数逼近0l范数的效果要比用2l范数逼近的效果好,但当训练样本较少且各类样本不平衡时,性能欠佳。此外,1l范数逼近稀疏表示求解涉及大量迭代,计算更加复杂、耗时较长。为解决训练数据有限性、不平衡和计算复杂度高等问题,本论文在协作表示分类器(CRC)纳入所有类样本字典协作表示参与分类的基础上,提出修正协作表示分类器(MCRC)。MCRC首先将由原始训练样本构成的字典通过奇异值分解映射成各类平衡的元样本字典,然后依据对测试样本的线性表示建立协作表示模型,同时施加KL散度(KLD)距离加权矩阵并约束与特征值相关的偏差加权和,实现适度加大两类残差距离的目标,改善分类性能。基于国内外学术界认可的AVEC2013抑郁症语料库的朗读语音部分完成性能评价实验。将语料库语句按长度差异分割成三种实验数据集,在留一交叉验证框架下进行MDD检测,对比SVM、SRC、标准CRC和MCRC的检测性能。结果表明MCRC算法的精确度(Accuracy)最高,且灵敏度(Sensitivity)与特异度(Specificity)更为匹配。此外,因MCRC仍能预先计算投影矩阵,大大降低计算开销。本文主要创新工作如下:(1)将协作表示思想首次应用于语音抑郁症检测。(2)提出修正协作表示检测模型。其优点:1)适用于小样本、不平衡样本的分类问题;2)分类器比较稳定,对正则化参数变化不敏感;3)受所处理语音段长度变化影响很小。
[Abstract]:Severe depression (MDD) is a common mental disorder. Patients have long term excessive sadness, negative emotion and cognitive impairment and even suicide, which seriously affect their physical and mental health. In the second edition of Beck depression scale, the range of MDD score is a sign that the phonetics of 29~63.MDD patients are usually monotonous, low and inactive. There are essential differences in the phonetic source, system and prosodic information contained in depression. Existing studies have shown that MDD. can be effectively detected by capturing acoustic characteristics. However, the limited and uneven distribution of depression corpus has become the main obstacle for severe depression detection. Support vector machine (SVM), Gauss mixed model (GMM) and sparse representation classifier (SRC). When all kinds of training samples are sufficient, GMM performance is good.SVM although suitable for small sample classification problem, but the number of samples should be balanced. When the sample is enough, the effect of sparse representation to the 0l norm with the 1L norm is better than the approximation of the 2l norm. The performance is good, but when the training samples are less and the various samples are not balanced, the performance is poor. In addition, the 1L norm approximation sparse representation involves a large number of iterations, and the computation is more complex and time-consuming. In order to solve the problem of limited, unbalanced and computational complexity of training data, this paper introduces the cooperative representation classifier (CRC) into all class sample dictionaries in this paper. On the basis of cooperative representation, a modified cooperative representation classifier (MCRC).MCRC is proposed to map the dictionaries composed of original training samples by singular value decomposition into all kinds of balanced meta sample dictionaries, and then a cooperative representation model is established based on the linear representation of the test samples, and the KL divergence (KLD) distance weighting matrix is applied at the same time. And constrain the deviation weighted sum associated with the eigenvalues, to achieve a moderate increase in the target of two kinds of residual distance and improve the classification performance. Based on the reading speech part of the AVEC2013 depression corpus recognized by the domestic and foreign academics, the performance evaluation experiment is completed. The corpus is divided into three experimental data sets according to the length difference, and a cross test is used. MDD detection under the certificate framework compares the detection performance of SVM, SRC, standard CRC and MCRC. The results show that the accuracy of the MCRC algorithm (Accuracy) is the highest, and the sensitivity (Sensitivity) is more matched with the specificity (Specificity). In addition, the projection matrix can be calculated in advance because MCRC can still be calculated in advance. The main innovations of this paper are as follows: (1) will cooperate with each other. The representation idea is first applied to speech depression detection. (2) a modified cooperative representation detection model is proposed. Its advantages are: 1) suitable for small sample, unbalanced sample classification problem; 2) the classifier is more stable, not sensitive to the regularization parameter change; 3) is affected by the change of the length of the speech segment.

【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R749.4;TN912.3

【相似文献】