Speech emotion recognition and depression prediction using deep neural networks

Abstract

Speech signals contain ample information from which computers can gain insight into the user's state, including emotion recognition and depression prediction. The applications are numerous, from customer service to suicide prevention due to depression. In this research, we propose several deep-learning-based methodologies to detect emotion, as well as depression. Deep neural networks variations such as deep feedforward networks and convolutional networks were used. The deep learning model training, multi-languages emotion and depression database have been utilized, using well-known databases such as the Berlin Emotion Database and DAIC-WOZ Depression Dataset. For speech emotion recognition, the algorithm yields an accuracy of 80.5% across 4 languages, English, German, French and Italian. For depression detection, the current algorithm obtains an accuracy of 60.1% tested with the DAIC-WOZ dataset. This research has also created the Sorrow Analysis Dataset – an English depression audio dataset that contains 64 individuals samples of depressed and not-depressed. Further testing achieved an average accuracy of 97% with 5-fold validation using 1-dimensional convolutional networks. Finally, a prototype currently in development with Skymind Xpress.ai is presented, outlining the design and possible applications in the real world. It has been shown that the model is capable of performing both training and inference on a Raspberry Pi 3B+.