Research on Depression Recognition Using Machine Learning from Speech

The lack of objective measures causes the most treatable mental illness, depression to be often under-diagnosed. Recent studies have shown that speech is a good indicator of depression, giving us a motivation to perform depression diagnosis using speech to create an objective measure. This paper studies the use of state-of-the-art machine learning (ML) models including ensemble in predicting depression severity using audio features after optimizing the data. We collect the audio data from Northwest Normal University and the classic paradigm of psychological self-reference effect (SRE) is used. In the task of vocabulary reading, the SVM model achieved very good performance. In the task of image description, the DT model was better than other models, and the F1 score reached 0.71. Our work has successfully shown that Random Forest (RF) trained using the mean of Zero-crossing Rate, Energy, Entropy of Energy, Spectral Centroid, Spectral Spread is a good model for depression prediction, which can predict the Beck Depression Inventory (BDI-II) at mean F1 of 0.71 and Root Mean Square Error (RMSE) of 4.21. The results are better than other state-of-the-art models. In the future, we aim to further verify the model's correctness and create an autonomous agent that could help depressed patients.