A Study on Speech Emotion Recognition Using a Deep Neural Network

Abstract

When using voice signals as input to a deep learning network, there may be myriad features depending on the method and purpose of extracting the voice signal features. Therefore, extraction of appropriate features should be conducted. In this study, verbal features necessary for speech emotion recognition (SER) and preprocessing features for a deep neural network are described in detail. We implemented various preprocessing methods using voice features. Also, a Keras-based deep neural network using Python libraries was implemented. With these features, we could obtain a test accuracy of 68.5 % using the deep neural network (DNN). As a result, we confirmed that the proposed DNN improved an accuracy by 30.1 % compared to a support vector machine (SVM).