Speech and Text-based Depression Detection On Social Media For Early Intervention

Depression is a significant mental health and SDG3 issue with far-reaching social and economic impacts. While early detection and intervention are crucial, traditional screening methods are often not accessible or accurate. This paper presents a method for detecting depression through a combination of voice and text analysis. By using log mel spectrograms and convolutional neural networks (CNNs) for voice analysis, and bi-directional long short-term memory (BiLSTM) networks for text analysis, the model improves detection accuracy and provides insights into mental health and well-being. The model structure, data collection, processing techniques, and training processes for both CNN and Bi-LSTM models are detailed. This study employs the "Suicide and Depression Detection" dataset from Kaggle for text and the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) for voice. A user interface is developed to ensure system accessibility. The experimental results demonstrate the high accuracy of the proposed approach for depression detection. This research highlights the importance of an accessible interface that combines multiple analysis methods and the need for interdisciplinary collaboration to address complex societal and mental health issues.