Summary of the Silent Speech Decoding Challenge Using EEG Data for Arabic Inner Speech Recognition

Abstract

The Silent Speech Decoding Challenge, hosted as part of the ABC 2025 conference, was designed to advance the decoding of silent (inner) speech from electroencephalography (EEG) signals, specifically targeting Arabic language processing. Participants leveraged advanced machine learning (ML) and deep learning (DL) techniques to classify EEG data from a uniquely collected dataset. This dataset, gathered by Qatar University’s Machine Learning Group, consisted of EEG recordings from ten subjects across multiple sessions, capturing both silent and overt speech for six distinct inner speech commands. Participants were tasked with developing robust ML/DL models to accurately decode these neural signals, addressing challenges such as generalization across subjects and robustness to noise and motion artifacts. The evaluation protocol included two tests—one focused on subject-wise accuracy and another emphasizing cross-subject generalization. Outcomes demonstrated significant strides in EEG signal processing, with top-performing models showing high accuracy, effective generalization, and noise resilience. This initiative notably impacts assistive communication technologies, potentially aiding individuals with speech impairments. Looking forward, future work involves expanding generalization capabilities, improving computational efficiency, and integrating these technologies into real-world applications. Overall, this challenge represents a crucial step toward innovative, accessible neural-based communication solutions.