Empowering Silent Voices: Adaptive EEG-to-Text Systems for Assistive Technologies

Abstract

Translating brain activity into natural language represents a transformative frontier in human-machine interaction and assistive communication technology for individuals with speech impairments. While electroencephalography (EEG) has shown promise for neural decoding, existing EEG-to-text methods remain constrained by closed vocabularies, limited semantic expressiveness, and inadequate accommodation of intersubject neural variability. This work presents a novel framework that transcends traditional closed-vocabulary limitations by synergistically combining subject-adaptive representation learning with advanced natural language processing architectures. Our approach employs deep neural networks to extract discriminative EEG features, enabling the generation of complex sentences that extend beyond the constraints of training data. Experimental evaluation on the ZuCo corpus demonstrates substantial improvements across multiple metrics, including BLEU, ROUGE, and BERTScore, surpassing state-of-the-art baselines. The framework effectively produces semantically coherent and grammatically accurate text while adapting to individual neural signal patterns through personalized modeling. By bridging open-vocabulary text generation with neural signal interpretation, this research establishes foundations for practical brain-to-text communication systems. The interdisciplinary implications span assistive technology innovation and personalized communication interfaces, advancing the paradigm of brain-computer interaction across clinical, research, and consumer applications.