Research On The Construction And Analysis Of Depression Corpus

Depression,also known as depressive disorder,is a serious mental disorder.It is clinically manifested as significant and lasting depression,decreased pleasure,reduced speech activities,etc.It has a suicidal tendency,which seriously affects people's physical and mental health,and also brings great harm to the society.As of 2017,300 million people worldwide suffer from depression,and there are more than 54 million people suffering from depression in my country.At present,the diagnosis method of depression is mainly subjective scale,which relies on the clinical experience of doctors and the degree of cooperation of patients,and lacks objective indicators.Therefore,as a non-intrusive,easy-to-collect,and low-cost objective indicator,speech has been favored by researchers.As we all know,the quality of the corpus directly affects the performance of the recognition model,and the same is true for the research on depression recognition based on speech signals.As far as we know,there are currently many publicly available depression corpora in foreign countries,but there is no publicly available depression corpus in China,and due to cultural and language differences,the results of foreign depression corpus research don't completely conform to our actual situation.Although the study of acoustic features is also an important part of the research on depressive speech,there are still inconsistent research results.At present,there are still no acoustic features of speech that can fully and effectively identify depression.Moreover,due to the differences in corpus data,classification system and network structure,the results can't be compared effectively.Based on this,this thesis first established a Chinese depression speech corpus,and secondly,extracted relevant acoustic features based on previous studies,and used experimental statistics to make a more detailed analysis and comparison of the acoustic features of depression and healthy subjects.The effective acoustic features and their changing laws are presented.Finally,this thesis uses the obtained effective acoustic features to perform modeling and recognition to prove the effectiveness of the establishment of the corpus.The main work and innovations of this thesis are as follows:1.Established a Chinese depression speech corpus.Based on the psychological self-processing abnormality theory,this thesis proposes a method of collecting depression corpus based on the classic experimental paradigm of psychology.Using different speech styles and emotional stimuli,combined with the classic experimental paradigm of psychology(Self-referential processing paradigm,autobiographical memory paradigm)to design depression text corpus.Finally,the speech of 60 subjects(30 depressed subjects and 30 healthy subjects)were collected,and the total duration was about 60 hours.The gender,age and educational background of the two groups of subjects were all matched.2.By using the method of experimental statistics,the extracted acoustic features are compared and analyzed.From the perspectives of speech style,emotional stimulation,and self-referential processing,the acoustic features that are significant under the above three angles were screened out as effective acoustic features for subjects of different genders.The effective acoustic features of female are loudness,MFCC3,F1 frequency and spectral flux,while those of male are MFCC1,MFCC5,fundamental envelope and F3 bandwidth.The correlation between the effective acoustic features and the scores of the Beck Depression Inventory was compared,and the analysis results showed that there is a certain correlation between the acoustic features and the severity of depression.3.Based on the corpus established in this thesis,depression recognition was carried out.This thesis uses Support Vector Machines to classify depression.From the perspective of vocalization methods,the classification rates of free speech,negative speech and self-perspective speech are respectively higher than other types of speech;the overall average classification rate of female speech is higher than the average classification rate of male speech.From the perspective of acoustic features,the classification rate of loudness and spectral flux of female is better,and the classification rate is more than 80%;the classification rate of fundamental envelope and MFCC1 of male is better,and the classification rate of fundamental frequency envelope is more than 75%,and the classification rate of MFCC1 is more than 80%.The above results also verify the validity of the corpus and the effective acoustic features.This thesis studies the expansion of depressive speech data,the selection of acoustic features,and the classification of depression,which will provide references for the future research on depressive speech.