EEG-to-Speech Decoding for Paralysis Rehabilitation Using Deep Learning Architectures

Abstract

Restoring communication for paralyzed individuals through EEG-based speech decoding remains a critical challenge in brain-computer interface (BCI) research. This study evaluates the performance of deep learning and traditional machine learning models in reconstructing speech from EEG signals. Unlike prior approaches that focus on a single frequency band, this study investigates the impact of multiple bands (Delta, Theta, Alpha, Beta, and Gamma) to comprehensively assess their contributions to imagined speech decoding. Additionally, we experiment with seven different EEG channel selection strategies, including full-channel analysis and selections derived from six key studies. Our methodology prioritizes temporal feature extraction, leveraging time-domain dynamics to enhance decoding accuracy. Using the 2020 International BCI Competition dataset, we compare four deep learning models: LSTM-RNN, CNN-1D, MLP, and MLP with Attention. While we include traditional models (e.g., KNN, CNN) from previous literature for contextual benchmarking, these models were not implemented in this study. Our findings highlight the significance of multi-band analysis and channel selection in improving EEG-based speech classification. This work advances the development of non-invasive, portable BCIs tailored for paralyzed populations, emphasizing practical deployability and real-time therapeutic potential.