We investigated the challenges of using noninvasive EEG-based brain-machine interfaces (BMIs) to decode imagined speech as a communication aid for individuals with severe physical disabilities. Our focus was on the limited representation of phonological features in the signal analysis related to speech articulation. To overcome this challenge, we examined an open-source speech imagery dataset to analyze spectral features and EEG channel covariance, and to assess their correlation with phonetics. Our logistic regression model distinguished phoneme classifications and revealed that EEG signals more accurately distinguish distinct words (such as ‘pat’ from /iy/) than similar sounding phonemes (such as /iy/ vs. /tiy/ or ‘pat’ vs. ‘pot’). Statistical tests showed significantly higher accuracy for “/iy/ vs. pot” over other comparisons, with no notable differences between “/iy/ vs. /tiy/” and “pat vs. pot”. The significant identified EEG characteristics were mainly low-frequency oscillations (4–30 Hz) in the brain’s frontal, temporal, and parietal regions, involving both hemispheres. This supports recent findings on the significant role of the right hemisphere in speech processing. Our findings indicate that surface EEG can detect neural patterns corresponding to specific phonemes, enhancing BMI performance. Future research should further explore phonological features to develop a comprehensive neural vocabulary for improved BMI operations.