Advancing Arabic Inner Speech Recognition with Machine Learning and Deep Learning

Abstract

Individuals with motor and communication impairments struggle to access communication tools and fully participate in society. Advances in Brain-Computer Interface (BCI) technology enable direct brain communication via noninvasive EEG signals from inner speech. However, accessible applications for Arabic-speaking individuals remain limited. Our research explores various feature extraction and classification methods on the publicly available Arabic EEG dataset, "ArEEG." Using EEGNet on raw EEG data, we improved cross-validation accuracy by 2%, reaching 27.1% per subject. Additionally, we propose a novel K-Nearest Neighbors (KNN) approach with Relative Wavelet Energy (RWE) and Gabor Transform features, selected via ANOVA, achieving 31.03% accuracy in subject-dependent analysis.