<p>Electroencephalogram (EEG) based Brain computer interface (BCI) emerges as a transformative technology with vast applications in neuroscience and rehabilitation. Imagined speech is the mental process of thinking and formulating words without vocalizing them through articulators. EEG signal is used to study imagined speech which can empower individuals with neurological impairments to communicate their thoughts effortlessly. The main challenge in decoding imagined speech is the nonstationary nature of EEG signals. Identifying robust features and scarcity of imagined speech datasets for properly training machine learning (ML) based algorithms is also a challenging task. The main objective of this study is to propose augmentation methods which mitigate data scarcity in EEG-based BCIs by introducing variations and strengthening model robustness through EEG data augmentation. The second objective is to propose a novel architecture capable of detecting variations in EEG signals for imagined speech datasets and show remarkable results. Seven diverse augmentation techniques are discussed, and the performance of the proposed model is analyzed in terms of accuracy, f1-score and kappa. The classification results are then compared with the case in which no data augmentation is used. The proposed model has shown remarkable accuracy of 91% for long words by using gaussian noise augmentation.</p>