Development of Sorrow Analysis Dataset for Speech Depression Prediction

Computers can get insight into the user's mental state, including depression prediction, by analyzing speech signals. Numerous uses exist, ranging from customer service to depression-related suicide prevention. In this study, we proposed a novel depression detection method based on deep learning. Deep neural network variants, 1D-CNN, 2D-CNN, and BiLSTM, were utilized. This research developed a new speech depression dataset, namely the Sorrow Analysis Dataset. It is an English depression audio dataset of 64 recordings of depressed and non-depressed individuals. Results showed that of the various architectures tested, 1D-CNN was found to produce the highest average accuracy of 97% with 5-fold validation.