| Decoding inner speech remains a complex challenge for non-invasive brain-computer interfaces (BCIs). This study presents a multimodal approach to classifying imagined speech by integrating simultaneous Electroencephalography (EEG) and functional Magnetic Resonance Imaging (fMRI) data. While EEG captures rapid neural oscillations, fMRI provides detailed spatial information about brain activity. We propose a robust pipeline incorporating artifact removal, deep learning-based feature extraction via EEGNet, and General Linear Model (GLM) analysis. A key contribution of this work is the systematic evaluation of dimensionality reduction strategies; we demonstrate that Analysis of Variance (ANOVA) feature selection followed by Principal Component Analysis (PCA) significantly enhances feature discriminability compared to standard techniques. Our results with an early fusion framework validate the complementary strength of fusing temporal and spatial modalities. We conclude by discussing the need for personalized, subject-specific models to address inter-individual variability, paving the way for more naturalistic thought-driven speech synthesis. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.