AdaBoost-Collaborative Representation Algorithm For Speech Severe Major Depressive Disorders Detection

In 2017,the World Health Organization(WHO)announced that 322 million people suffered from depression worldwide which was the second leading cause of death among15–29 year olds globally in 2015.Severe major depressive disorders(SMDD)whose score range is 29–63 in the Beck Depression Inventory-II will have serious self-mutilation,even suicidal tendencies,and become a potential high-risk group to trigger the current situation.Domestic and foreign scholars have made fruitful attempts and explorations in detecting and recognizing depression from the perspective of the impact of depression on phonological prosody and acoustic linguistic characteristics.However,speech pattern detection for SMDD needs further study.Although the speech of SMDD patients is often characterized by slow speech speed,long pause and single words,the data samples are less and the distribution of SMDD and non-SMDD is significantly imbalanced,which aggravate the dilemma of low difference of severity and render this problem challenging.Collaborative representation classifier relieves the limitation of training samples’ insufficiency to some extent by all of training samples participating in representing a test sample.However,there are too few samples of the target class to adequately describe their spatial distribution for the task of SMDD detection.Existing research results show that the performance of traditional depression detection methods based on single classifier has become a bottleneck,and generalization error is difficult to further reduce.Aiming to deal with these problems,this thesis adopts the Adaboost framework and proposes the detection model of AdaBoost-collaborative representation classifier(AdaBoost-CRC),Improve the accuracy of the classifier by weighting the base classifier.The main work of this paper: The algorithm uses the mean vector of frame-based Mel-frequency cepstral coefficients(MFCCs)as the utterance feature;Calculate the singular value decomposition of training samples with imbalanced class distribution to obtain metasamples and build a balanced class dictionary;Establish the ensemble detection model from the the balanced class dictionary.In the stage of feature extraction,the effect of feature perturbation at the utterance level is investigated by randomly selecting the threshold of voiced frames in a specified range.AdaBoost-CRC framework incorporates the number of randomly the metasamples of each class as dictionary atoms.By creating a variety of individual difference-based classifiers,a stochastic dynamic ensemble weighted classification model is constructed to enhance the heterogeneity of the ensemble detection system.The algorithm was evaluated using the AVEC2013 depression corpus.The SMDD detection performance of neural network,support vector machine,sparse representation classifier,collaborative representation classifier was compared following the strategy of theleave-one-speaker-out cross validation.The experimental results show that the proposed method is the best.