2-level hierarchical depression recognition method based on task-stimulated and integrated speech features

• A hierarchical classification model was designed considering the task-stimulated features and integrated features for better recognition performance. • I-vector was used to solve the variable length problem of frame level features and overcome speaker and channel variability effects. • The effectiveness of hierarchical classification was verified on different features and their combinations. • Gender-independent and gender-dependent experiments were carried out to test the gender influence on our method. Depression had been paid more and more attention by researchers because of its high prevalence, recurrence, disability and mortality. Speech depression recognition had become a research hotspot due to its advantages of non-invasiveness and easy access to data. However, the problems such as the speech variation in different emotional stimulus, gender impact, the speaker and channel variation and the variable length of frame feature, would have a great impact on recognition performance. In order to solve these problems, a novel 2-level hierarchical depression recognition method was proposed in this paper. It contained two stages. In 1 st -level classification stage, i-vectors were extracted based on spectral features, prosodic features, formants and voice quality of speech segments in different task stimulus respectively. Then, support vector machine (SVM) and random forest (RF) were used to obtain primary results. In the stage of 2 nd -level classification, the results of tasks with significant accuracy differences were aggregated into new integrated features. The final result was achieved on new features by SVM. Our experiments were based on the depression speech database of the Gansu Provincial Key Laboratory of Wearable Computing. The experimental results showed that the proposed method had achieved good results in both gender-independent and gender-dependent experiments. Compared with baseline method and bagging classification, the highest accuracy of our method was raised by 9.62% and 9.49% respectively in gender-independent experiments, and F1 score also got improvement obviously. The results also showed that our method had better robustness on gender effect.