A Novel Approach to Neural Network Based Intelligent Systems for Emotion Recognition in Audio Signals

Abstract

In this work, we present research done in the field of affective computing and intelligent systems which could benefit countless other fields including medicine, game development, and robotics. We attempted to develop an intelligent system capable of recognizing seven basic human emotions when given nothing but a raw audio signal. This task was accomplished through creating a two-tier intelligent system comprised of neural networks designed to determine the emotion of a 40 millisecond audio signal and a simple voting algorithm used to combine the output of the neural networks. Overall, we were able to achieve 60% accuracy of classifying 7 different emotions in a semi-open loop test with the system's primary guess, and 80% accuracy with the addition of the secondary guess. The results we obtained prove the potential of using this unique system design in the field of affective computing and hint that greater accuracy could be achieved through the combination of multiple intelligent systems.