Improvement of Speech Emotion Recognition by Deep Convolutional Neural Network and Speech Features

Speech emotion recognition (SER) is a dynamic area of research which includes features extraction, classification and adaptation of speech emotion dataset. There are many applications where human emotions play a vital role for giving smart solutions. Some of these applications are vehicle communications, classification of satisfied and unsatisfied customers in call centers, in-car board system based on information on drivers' mental state, human-computer interaction system and others. In this contribution, an improved emotion recognition technique has been proposed with Deep Convolutional Neural Network (DCNN) by using both speech spectral and prosodic features to classify seven human emotions-anger, disgust, fear, happiness, neutral, sadness and surprise. The proposed idea is implemented on different datasets such as RAVDESS, SAVEE, TESS and CREMA-D with accuracy of 96.54%, 92.38%, 99.42% and 87.90%, respectively, and compared with other predefined machine learning and deep learning methods. To test the real-time accuracy of the model, it has been implemented on the combined datasets with accuracy of 90.27%. This research can be useful for development of smart applications in mobile devices, household robots and online learning management system.