Speech-based Sentiment Recognition System using PDCNN and LSTM Algorithms

Abstract

Sentiment analysis is very important for the multiple human-computer interaction system. Many deep learning frameworks have been presented for sentiment analysis using speech signals. However, the performance of speech-based sentiment recognition is limited because of poor generalization capability, limited long-term dependency, inferior feature representation, poor balance in speech spectral and temporal properties, and complexity in deep learning algorithm frameworks. This paper presents speech-based sentiment recognition (SSR) using a parallel deep convolution neural network, a long short-term memory (DCNN-LSTM) network, and multiple acoustic features (MAF). The multiple acoustic features consist of spectral, time domain, and voice quality features for improving the feature distinctiveness of the speech signal. Further, the Archimedes Optimization algorithm (AoA) selects the prominent MAFs. The weights of the fitness function of the AoA algorithms are automatically optimized using the Multi-Attribute Criteria Theory (MAUT) algorithm. The outcomes of the proposed algorithm are evaluated on the Berlin Sentiment Database (EMODB), which consists of seven sentiments: happiness, anger, boredom, disgust, fear, neutral, and sadness.