Enhancing Silent Speech Decoding Using EEG and Gradient Boosting Classifiers: A Comparative Analysis of XGBoost and LightGBM on a Native Arabic Silent Speech Dataset

Silent speech decoding using EEG signals is a key area in Brain-Computer Interfaces (BCIs), enabling communication without vocalized speech. This study is part of the Silent Speech Decoding Challenge (SSDC) and focuses on classifying six imagined speech commands: Up, Down, Left, Right, Select, and Cancel.EEG signals from eight subjects underwent preprocessing, including bandpass (0.5-50 Hz) and notch filtering, Independent Component Analysis (ICA) for artifact removal and segmentation using a 250 ms sliding window with $50 \%$ overlap. Time-domain features were extracted and normalized before classification.Models-Naïve Bayes, KNN, SVM, XGBoost, and Light-GBM-were evaluated using 10-fold cross-validation. While baseline models showed moderate performance (SVM: 24.3% accuracy, KNN: 35.2%), XGBoost and LightGBM achieved perfect validation accuracy (100%). An independent test set from two subjects was also analyzed, though ground truth labels were unavailable for objective assessment.These results highlight the effectiveness of hand-crafted features with gradient-boosting classifiers for silent speech recognition, demonstrating the potential of EEG-based BCIs for real-world applications.