What question did this study set out to answer?

This research aims to improve imagined speech recognition using EEG data through robust feature analysis.

May 7, 2026International Journal of Advanced Computer Science and Applications0 citationsOpen Access

Feature-Level Analysis and Robust Baselines for EEG-Based Imagined Speech Recognition on the ASU Dataset

HDHatem T M Duhair MIMasrullizam Mat IbrahimTechnical University of Malaysia Malacca JAJamil Abedalrahim Jamil AlsayaydehTechnical University of Malaysia Malacca

Key Points

This research aims to improve imagined speech recognition using EEG data through robust feature analysis.
Utilized the ASU imagined speech dataset with ten target classes.
Preprocessed 6,520 EEG trials with 60 channels using bandpass filtering and z-score normalization.
Constructed a comprehensive feature representation integrating multiple EEG-derived metrics.
Multinomial logistic regression achieved approximately 0.41 accuracy and 0.42 macro F1 on the ten-class task.
Identified autocorrelation and power spectral density as key features in the predictive model.
Deeper models like CNN–LSTM did not surpass logistic regression performance, indicating overfitting.

Abstract

Imagined speech decoding from non-invasive electroencephalography remains a challenging problem, especially when moving beyond small vocabularies and optimistic evaluation protocols. This work revisits the Arizona State University (ASU) imagined speech dataset and treats it as a rigorous ten-class benchmark, with a focus on offline, corpus-level analysis rather than real-time deployment. After unifying all recordings into 5 s epochs at 256 Hz, 6,520 trials with 60 EEG channels were preprocessed using bandpass filtering, baseline correction, z-score normalization, and trial-wise ICA for artifact attenuation. On top of this pipeline, a comprehensive feature representation was constructed that combines common spatial patterns, discrete wavelet statistics, time-domain moments, autocorrelation coefficients, power spectral density band powers, and Hjorth parameters into a single 5,120-dimensional vector. A block-wise ablation indicates that autocorrelation, CSP, PSD, and Hjorth features carry most of the discriminative information in this setting, while wavelet and simple statistical descriptors contribute little and can be removed without harming performance. Using only the informative blocks (3,440 features), a multinomial logistic regression classifier reaches about 0.41 accuracy and 0.42 macro F1 on the ten-class task, roughly four times chance level. A multi-layer perceptron and a CNN–LSTM model, trained under the same splits and with class weighting, do not outperform this linear baseline and exhibit stronger overfitting. Within the evaluated protocol, these findings suggest that carefully engineered features capture most of the discriminative structure accessible on this corpus, and that deeper models add complexity without clear benefit. The study provides a transparent baseline and a feature-level analysis that can serve as a reference point for future work on imagined speech recognition and transfer learning across EEG corpora.

Ask AI

Helpful

Bookmark

View Full Paper