A Novel Approach for Depression Detection Using Audio Sentiment Analysis

Depression is a medical illness that affects an individual negatively by changing the way one feels, think and act but luckily, it's treatable, so the problem is its detection which we can solve with the help of machine learning. A model is developed to detect whether a person is suffering from depression or not using the prosodic features (pitch, tone, rhythm) of their voice which are promising indicators of depression. The model is trained by the data provided by the DAIC-WOZ that contains clinical interviews taken by the virtual interviewer called Ellie, which is then preprocessed using sox and system programming to remove the voice of virtual interviewer. Feature extraction is done by making spectrograms of each audio file using Librosa library. Then these spectrograms are passed into a convolution neural network with average pooling layers, dropout, He-initialisation, batch normalization, exponential linear unit activation function, and Nesterov accelerated gradients optimizer.It has been observed from the experiments that the prediction is made with an average F1 score of 0.93.