Solving emotion classification problem using deep learning

Abstract

Speech emotion classification is one of the most interesting and complicated problems in to-day’s world. One of the main obstacles to this task is that emotions are subjective and difficult to capture. In this paper, we proposed deep learning methods that solve emotion classification problems based on audio streams. Three methods are propagated and compared throughout the paper. Within the first method a Mul-tilayer Perceptron model was built. A second method shows decreased accuracy building Long Short Term Memory models. Finally, the third method that reached the best accuracy among others is convolutional neural network models. A speech corpus consisting of acted and spontaneous emotion samples in English language is described in detail. This dataset was tested and trained using these proposed methods. The CNN model for our emotion classification problem achieved a validation accuracy of 70%.