Deep Learning based Speech Emotion Recognition using Multiple Acoustic Features

Automatic speech emotion recognition (ASER) is crucial for multiple robotics applications, social media platforms, human-machine interaction systems, auto-chatbots, etc. However, the ASER performance is poor due to inadequate representation of the emotional speech, poor feature distinctiveness, and less generalization capability. This paper presents ASER based on a Deep Convolutional Neural Network (DCNN) that accepts the multiple acoustic features (MAF) to describe emotional speech. Further, it uses a combination of hybrid particle swarm optimization and Archimedes optimization algorithm (PSO-AOA) for selecting the essential features from MAFs. The outcomes of the MAF-PSO-AOA-DCNN are assessed on the BAUM dataset. It provides an overall accuracy of 95.20%, precision of 0.96, recall of 0.94, and F1-score of 0.95, which is superior to traditional techniques.