Frequency-embedded sequence learning for multi-class EEG-to-text interface

A brain-computer interface (BCI) is an emerging technology for decoding users' intentions, enabling speech decoding from electroencephalography (EEG) signals. This study proposes a multi-class (i.e., 5 words or phrases) EEG-to-Text classification pipeline based on a lightweight 1−dimensional convolutional neural network (1D−CNN) architecture. The model is trained using both time-domain EEG signals and frequency-embedded power spectral density (PSD) sequences, facilitating systematic comparison of feature representations. EEG signals were preprocessed using a bandpass filter (4 − 120 Hz) and a notch filter (60 Hz) followed by PSD estimation using the modified Welch method. The processed signals were then used to train the proposed 1D−CNN architecture designed for sequential data. In intra-subject evaluation, the proposed approach achieves classification accuracies (mean ± standard deviation) of 46.1 ± 5.1 and 64.4 ± 8.3 using time-domain EEG signals and frequency-embedded PSD sequences, respectively. Peak intra-subject accuracies reach 54% and 82% using EEG signals and PSD sequences, respectively. For inter-subject (pairwise) evaluation, the best observed accuracy is 42%. Inter-subject (pooled) performance remains near chance level (23.3 ± 1.0 and 24.5 ± 0.4), highlighting inter-subject generalization as a key challenge in EEG-based BCI systems. These results demonstrate that while intra-subject performance is promising, inter-subject generalization remains a major limitation. Overall, the proposed approach demonstrates the potential of frequency-embedded PSD representations for improving EEG-based imagined speech classification, while motivating further work toward robust, user-independent BCI systems.