The aim of this study by Meenakshi Bisla and Radhey Shyam Anand was to address the gaps involved in interpreting imagined speech using EEG signals and classify these signals into distinct categories. These categories include long word (‘cooperate’ and ‘independent’), short word (‘in ’, ‘out’, and ‘up’), vowel (‘/a/’, ‘/i/’, ‘/u/’), and short-long word (‘in’ and ‘cooperate’) imaginations. In this study, 15 subjects’ (11 male and 4 female) EEG signals were recorded for each of the one hundred word imaginations that belonged to each of the four categories listed above. The data set was enlarged via sliding window augmentation, which Bisla and Anand claim to invite variability, improve performance, and reduce overfitting. After analysis, the top fifty features of the data were extracted, which were then input to the machine learning (ML) classifiers for imagined speech interpretation. This study implemented six ML classifiers of EEG data: SVM, RF, KNN, XGBoost, LightGBM, and CatBoost. The Grind Search method was used to tune the parameters of each classifier to optimize each one’s performance. To maximize efficiency, much of the data was divided into subsets before interpretation. The reason for this nesting is to prevent overfitting and any unintentional bias introduced in larger data sets. There were a few different classification approaches from the ML model, including a subject-dependent method, an all-but-one method, and a subject-independent method, where the amount of information coming directly from research participants varies with each method. All three of these were applied to each of the six models. The result of this experiment was that the methods reached average accuracies of 78.52% for short words and 80.21% for vowel classification, and the CatBoost classifier had the highest accuracy overall. The best performance stemmed from using short-long words, and it was found that using the top 25 features was more accurate than using the extracted fifty.