Research On Depression Recognition Based On Speech Signal And Samples Sparse Approximate

In recent years,many results have been obtained in the study of using physiology and behavior signals such as EEG,speech,facial expression,eye movement,and gait to objectively detect depression.Speech signal has become one of the research hotspots in this field due to its advantages of easy access,non-intrusiveness,low cost,and large amount of information.In practice,the speech data of depressed patients is generally obtained in a multi-center(location)and multi-stimulus(question)manner,which leads to changes in the marginal distribution of the features of the samples and weak label problems,which in turn poses a challenge to the generalization ability of the recognition model.Aiming at the problems of multi-center and multi-stimulus,this paper effectively combines affine hull and sparse approximation to construct a depression recognition model with stronger generalization ability.The main contributions and innovations of this paper are as follows:1.Aiming at the problem that the features marginal distribution of samples in a multi-center scene changes and the generalization performance of the model is weakened,we proposed a depression recognition method based on affine hull and sparse approximated nearest points.This method uses a pair of sparse approximated nearest points between a set of depressed samples and non-depressed samples to construct a hyperplane for depression recognition.The cross-validation results on different center data sets show that the model can effectively reduce the prediction error in multi-center scenario.2.Aiming at the weak label problem of speech data in multi-stimulus scenario,we proposed a depressed speech set recognition method based on sparse approximated nearest points.This method takes the speech data of the same subject under multiple stimuli as a set,uses the sparse approximated distance to measure the similarity between sets,and uses the k-nearest neighbor method to discriminate label.The experimental results show that the model can effectively improve the weak label problem.It can achieve a recognition accuracy of 79.3%on 237 samples(135 depression,102 control),which is better than the multi-speech decision-layer fusion model based on weak label.This paper explores the speech-based depression recognition method in multi-center and multi-stimulus scenarios.Using the optimization theory and the sparse nature of the₁ norm,a depression recognition model is constructed from a new perspective,which effectively improves the generalization of the model in multi-center scenario and the weak label problem of speech data in multi-stimulus scenario.