Development of a Depression Detection System using Speech and Text Data

Depression is a prevalent mental health condition that impacts a considerable proportion of the world's population. However, early detection and intervention are crucial for effective treatment. Current depression detection models often rely on a single modality, either speech or text data, which may lead to inaccurate results due to the limited information being considered. In this paper, we propose a novel approach for the accurate detection of depression in individuals using both speech and text data. This approach addresses the limitations of existing solutions by combining objective measures of speech and text data, providing a non-intrusive and efficient means for depression detection. In this project, two models were trained: a speech emotion recognition model using Long Short-Term Memory (LSTM) and a text model based on the Random Forest Method. The TESS dataset was used to train the speech emotion recognition model, whereas the text model was trained on the Twitter dataset, which is accessible on Kaggle. The results indicate that our proposed approach is highly effective in detecting depression, with an accuracy of 98% achieved by the speech emotion recognition model and 95% by the depression detection model on the testing data. The outputs of both models are combined using a decision tree method, resulting in an accuracy of 100%. The proposed method employs a Decision Tree algorithm that takes the output of both the Speech Emotion Recognition (SER) model and the text-based depression detection model as inputs and applies a set of rules to classify users into one of three categories: depressed, mildly depressed, or not depressed. Then a webpage was developed where users can input their speech and text data and receive a prediction of their depression status based on the integrated output of both models. Overall, the results demonstrate that the proposed system provides an assuring solution for the early diagnosis of depressive symptoms.