Phoneme-based Imagined Vowel Identification from Electroencephalographic Sub-Band Oscillations during Speech Imagery Procedures

Speech Imagery (SI) corresponds to imagining speaking an intended speech or a segment of speech. Decoding the SI process aids in building speech-based neural prosthetic devices. Though SI-based research has been carried out to decode imagined speech for more than a decade, there is a lag in achieving the naturalness of the spoken language. This is because the words are built as the combination of phonemes in any natural language, but the research so far has been involving the SI of vowels only. Hence, this work focuses on identifying the vowels from EEG signals acquired while imagining the corresponding phonemes. The acquisition process was repeated for multiple trials. The EEG signals were decomposed into five sub-band frequencies to analyze the activity during SI tasks. The energy coefficients extracted from the sub-band frequencies were employed in training the Recurrent Neural Network to classify the English vowels. Further, to emphasize the importance of training the classifier with multi-trial data, the results were compared with that of the single-trial data acquired from the same set of participants, and an accuracy of 84.5% and 88.9% were achieved for single and multi-trial protocols, respectively. The analysis using multi-trial data was able to achieve 4.4% higher accuracy when compared to single-trial data. Higher activations in the theta band during the speech imagery tasks and higher Classification accuracy while applying theta band features show the capability of using the theta band features in imagined speech decoding tasks.