Emotion recognition in Arabic speech signal

Abstract

Recognizing emotions has become an area of great interest to researchers in the past few years. Emotion recognition is a multidisciplinary area, among which is the recognition of emotions from speech. Recognizing speech emotion is a significant endeavor in human speech processing and developing human-computer interaction. This work presents the performance of machine learning approaches for the recognition of emotions from an Arabic speech signal. Initially, we used the Lebanese audio database Arabic-Natural-Audio-Dataset (ANAD), which contains 384 records with 505 happy, 137 surprises, and 741 angry units. Next, we use the OpenSMILE toolkit to extract the necessary speech features with two methods, Low-Level Descriptors (LLDs) with 988 features, and Mel-frequency cepstral coefficient (MFCC) with 39 features. Also, we applied features selection on LLDs and MFCC using Learner Based Feature Selection. We suggested Rough set theory for select features in order to improve results. Then, for classifying the emotions into different classes, Multilayer Perceptron (MLP), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Logistic Regression (LR) are employed. Results showed that MLP outperformed other models when applied on LLDs and MFCC features with accuracy 87%, 83% respectively.