Imagined speech refers to the mental process of generating the sound of words without physical vocalization. Decoding brain signals associated with imagined speech has the potential to revolutionize communication for individuals with physical disabilities, such as those suffering from locked-in syndrome. Additionally, this technology can be applied in scenarios where verbal communication is challenging, such as in military operations. In this study, we utilize EEG signals to decode imagined speech, focusing on four classes of words from the FEIS dataset, which contains 16 classes of words and phonemes. Instead of integrating machine learning models, we propose an adaptive technique that separately employs three distinct algorithms: Support Vector Machine (SVM), Decision Tree, and Linear Discriminant Analysis (LDA). Our results indicate that while LDA demonstrated reasonable performance, the Decision Tree classifier exhibited stable accuracy even as the number of subjects increased. In contrast, the SVM achieved higher accuracy with fewer subjects but showed a significant decline in accuracy with more subjects. The SVM's effectiveness in capturing complex patterns from EEG signals underscores its potential yet highlights the need for balancing accuracy and scalability. This adaptive approach offers a flexible solution for EEG-based speech decoding, with significant potential for assistive communication technologies and other fields where non-verbal interaction is crucial.