Depression recognition base on acoustic speech model of Multi-task emotional stimulus

•We proposed MCL-mRMR feature selection algorithm to solve the problem of poor stability and low accuracy of feature set, which was caused by the influence of emotional stimuli on speech signals.•The possible explanation for the effectiveness of selected features was given, and the difference between the emotional valence related HSFs and depression related HSFs were found.•MTSW-Bagging was implemented, which made full use of the emotional stimulus information and improved the depression recognition performance.•Our contribution also included setting up multiple emotional stimulus tasks to collect subjects' voice data for testing proposed methods. Depression places great burden on families and society owning to its high prevalence recurrence and disability mortality. Using efficient and objective methods to recognized depression has attracted more and more attention from researchers. Subtle changes in the speaker's physical and mental state will be subconsciously reflected in vocal apparatus. Individuals have different responses to different emotional stimuli. Speech signals are easily affected by emotional stimuli, and thus will have a great impact on depression recognition. This study has two aims, first was to collect speech data in different emotional stimulus (positive, neutral and negative), and explore effective feature set with strong interpretability. The second aim was to design efficient multi-task recognition model. A depression recognition method based on max-relevance and min-redundancy (mRMR) with multi-class labels (MCL-mRMR) and multi-task stimulus weighted Bagging (MTSW-Bagging) classifier was proposed. Firstly, MCL-mRMR selected features which had high correlation with emotional valence and depression, meanwhile features' dimensions decreased. Next, MTSW-Bagging classifier was designed to recognize depression, whose base classifier was composed of weighted multi-task emotional stimulus classifiers. Experimental results showed that the features selected by MCL-mRMR had higher performance with the accuracy and F1 score were increased by 5.59% and 4.2% respectively compared with the original full features. Meanwhile, our proposed method was superior to baseline method with an improvement of 13.2% and 12.8% on accuracy and F1 score respectively. Compared with state-of-the-art related methods, our method also had its superiority of strong interpretability of features and being independent of training data scale.