Alphabet Imagined Speech Classification based on EEG Data using Network Transformer Spatio-Temporal

Abstract

Brain-Computer Interface (BCI) based on Electroencephalography (EEG) signals enables direct communication through brain signals. EEG is widely adopted due to its non-invasive nature and low cost. EEG-based BCI for imagined speech classification still faces significant challenges due to limited data and differences in brain activation patterns across cognitive stimulus types, such as letters, numbers, objects, and even subjects. This study proposes a cross-domain transfer learning approach using a Network Transformer Spatio-Temporal (NetTraST), Transformer-based model pre-trained on the Character dataset (10 classes of alphabet letters), and also the Continuous Learning technique. This study evaluates the initial generalization ability on the Digit (numbers) dataset, then performs fine-tuning to improve domain adaptation. The next stage, Final Boost, is performed with a partial-unfreeze layer, low learning rate, label smoothing, and Exponential Moving Average (EMA) techniques to stabilize weight updates. Meanwhile, on the Continuous Learning technique, we combined all data from three categories, implicates the number of classes increased from ten to thirty. Furthermore, this study conducts extensive classification experiments under various parameter settings and learning strategies to systematically evaluate model robustness and performance. Experimental results show that transfer learning and continuous learning achieve accuracy of 93.67% and 96.78%, respectively. This study fills a research gap that has not been explored in previous studies.