Effective Features Analysis And Application In Depression Level Evaluation Based On Speech Data

Depression is a common mental disorder with characteristics as high morbidity,recurrence rate,suicide rate and low awareness rate and treatment rate.It will bring great damage to individual's physical and mental health.In recent years,the morbidity rate of depression rose gradually with the constant increase of social stress.However,current diagnosis methods are limited to psychologist's observation and patients' self-description,leading to subjective bias in some ways.An objective,efficient and convenient assessment method is needed to assist diagnosis.Speech is a potential powerful indicator of depression with the characteristic of non-invasion and cheapness.According to research method,researches in the domain of depressed speech can be divided into cross-sectional study and tracing study roughly.In tracing studies,speech data of patients was collected regularly during treatment to observe the trend of speech features value along with the change of depression level,but such studies can only pay attention to individual's speech,the conclusions they drew may not be applicable in crowd classification;Cross-sectional scholars collected speech data in a short time interval.They focused on the differences in speech features between normal and depressed people,while fewer studies pay attention to the differences between patients with different depression level,and inconsistent conclusions were drawn due to subjects' individual factors.Unlike the essential psychology and physiology differences between depressed patients and normal controls,The differences between patients of different depression level are not obvious,leaving a challenge in detecting the speech distinction between them.Up to now,few effective speech feature was proposed.Contraposed the problem,and considered individual differences,the main works and contributions were described as follow:(1)A three-class speech dataset was constructed,and new features that haven't been discussed in pre-studies were added.132 subjects(72 females and 60 males contained)were recruited,and they were divided into normal,mild and severe depression according to PHQ-9 score.Their ages,education background and professions were well matched to eliminate the effects of interferential factors.Common paradigms and emotion stimuli were used to motivate speech,therefore the dataset was constructed.The dataset contained 14 categories of features,including classical features listed in pre-studies,and several features which have not been discussed.(2)Several methods were applied in analysis and selection of speech feature set,and several effective feature sets,including prosodic features like voice intensity and spectrum features like MFCCs and LPCs were found.In the three-class(contains classes of normal,mild depression and severe depression crowds)analysis of speech data,five and four feature sets were selected respectively from male and female data,and the classification accuracy in three classes of them all reached more than 60%.The feature sets were analyzed based on their physical property to explore the reason why they can distinguish crowds with different depression level effectively.(3)A multi-feature sets decision fusion model was built utilizing the feature sets mentioned above,and put into use in speech data to improve the performance of speech data in assessment of depression level.The predicting accuracy in speech data of male and female reached 70%?75%,higher than that of previous studies.Based on the speech data,we took advantage of several statistics and data dimensionality reduction methods to find several effective feature sets,which performed well in three-class analysis of speech data and assessment of depression level.This finding will provide a foundation in assessment of depression level in use of speech data.