Automatic speech emotion recognition for arabic dialects: a new dataset and machine learning framework

Article

Jan 01, 2026

DOI: 10.1007/s10586-025-05830-y

Zineddine Sarhani Kahhoul Nadjiba Terki Habiba Dahmani Belkacem Athamena Zina Houhamdi Madina Hamiane Mohammed Bourennane

Abstract

Automatic Speech Emotion Recognition (ASER) is a critical aspect of affective computing, which detects emotions in speech to facilitate efficient human-computer interaction. An area that has received little attention in previous research is the Algerian Arabic dialect, which is the setting in which this study examines ASER. We introduce a new corpus, Open Your Heart (OYH), which consists of roughly 6.3 hours of emotional spontaneous speech taken from a talk show on television. A wide variety of emotional expressions are captured in the 6,167 audio clips from 43 male and female speakers that make up the dataset. These expressions are categorized and analyzed through the Geneva Wheel of Emotions (GWE), providing an in-depth understanding of the emotional spectrum. We use the openSMILE toolkit to pull out audio features, then choose the best ones using the Backward Feature Elimination (BFE) method and an improved version that removes the least useful features to make the feature set better. We employ Support Vector Machines (SVM) as the primary classification model, alongside ten additional machine learning classifiers. We assess each model using different numbers of estimators. Among all classifiers, SVM achieves the highest performance, reaching a maximum accuracy of approximately 59% with a complexity of 0.00001. These findings surpass previous benchmarks, demonstrating the robustness of the proposed methodology for emotion recognition in Arabic speech.