Spectral Analysis of EEG Signals for Automatic Imagined Speech Recognition

Brain–computer interface (BCI) systems are intended to provide a means of communication for both the healthy and those suffering from neurological disorders. Imagined speech conveys users intentions. This article investigates the feasibility of spectral characteristics of the electroencephalogram (EEG) signals involved in imagined speech recognition. Eleven subjects were recruited to perform the speech imagination task. This article analyses the spectral features for binary and multiclass classification of imagined words in six different frequency bands (FBs). The 1-D EEG signals were converted into time–frequency representation (TFR) plots using smoothed pseudo-Wigner–Ville distribution (SPWVD) and classified using a convolutional neural network (CNN). In addition, the analysis was performed for subject-dependent, subject-independent, and leave-one-subject-out (LOSO) approaches along with the all-data approach. The proposed method achieved promising results in the gamma band with a binary classification accuracy of 82.04% ± 2.45%, 81.66% ± 4.93%, 78.97% ± 3.12%, and 81.04% ± 3.08% in all-data, subject-dependent, subject-independent, and LOSO approaches, respectively, and a multiclass classification accuracy of 51.44% ± 3.55%, 50.20% ± 1.35%, 49.93% ± 1.72%, and 50.42% ± 2.18% in all-data, subject-dependent, subject-independent, and LOSO approaches, respectively. Finally, the multiclass scalability in decoding the imagined words is investigated by increasing the number of classes from 2 to 15. The study’s findings demonstrate that the EEG-based imagined speech recognition using spectral analysis has the potential to be an effective tool for speech recognition in practical BCI applications. The contribution of this article lies in developing an EEG-based automatic imagined speech recognition (AISR) system that offers high accuracy and reliability while also providing a noninvasive method for speech recognition.