Abstract
Background Decoding inner speech from electroencephalogram signals presents a promising direction for advancing brain-computer interface technologies. However, this remains a highly challenging task because of factors such as the inherently low signal-to-noise ratio, significant inter-subject variability, and the complex, non-stationary nature of neural activity captured in EEG data. Methods This study presents a novel classification framework that combines instantaneous phase-locking value and common spatial patterns for feature extraction. The dimensionality of the instantaneous phase-locking features was reduced using principal component analysis. The two feature types are processed independently using Recurrent Neural Networks and Deep Neural Networks and then concatenated for the final classification. The model was evaluated under both subject-dependent (S-d) and subject-independent (S-Ind) settings on two publicly available EEG dataset. The first dataset includes imagined speech from five individuals in two categories: social and numerical, while the second dataset includes inner speech from ten participants performing four Spanish-word commands. Results The proposed model demonstrated strong performance in the S-d setting, achieving average accuracies of 95.16 ± 3.50% for social words and 95.96 ± 2.40% for numerical words, with macro F1-scores exceeding 95.19%. For the second dataset, the mean S-d accuracy was 90.23 ± 10.61%, while S-Ind accuracies were 64.09 ± 3.20% and 52.19 ± 2.64% for the second and first datasets, respectively, highlighting the challenge of cross-subject generalization in EEG-based inner speech decoding. Notably, the proposed method outperformed previously reported approaches on both datasets, achieving relative improvements of up to 226.21%. Conclusion The proposed approach shows strong performance in subject-dependent inner speech decoding and provides additional validation across two publicly available EEG datasets with different participants and task configurations. While the method shows practical potential, the results also highlight the ongoing challenge of generalizing across individuals, motivating further evaluation on larger, more diverse, and multi-session datasets.