Multiple Acoustic Feature-Based Speech Emotion Recognition for Sentiment Analysis

Sentiment analysis is critical in multiple affective systems encompassing robotics, chatbots, affective robots, call centers, audio conferences, etc. The sentiments affect the voice's spectral, temporal, and quality parameters. This paper analyses the effect of emotions on spectral, temporal, and voice quality characteristics of the voice. It provides sentiment classification using multiple acoustic features (MAF) comprising spectral domain (SDF), time domain (TDF), and voice quality features (VQF) of the voice. It uses different machine learning and deep learning classifiers for the sentiment classifications, such as K-nearest neighbor Classifier (KNN), Support Vector Machine (SVM), Classification Tree (CT), Random Forest (RF), and Deep Convolutional Neural Network (DCNN) based classifiers. The outcomes of the methods are estimated based on the accuracy of the EMODB dataset. The DCNN provides improved accuracy of 94.3% than RF (93.3%), CT (82.4%), Linear SVM (89.5%), RBF SVM (84.8%), and KNN (71.9%) for ASSR for EMODB dataset.