Emotions significantly influence human cognition and decision-making. Electroencephalography (EEG), with its high temporal resolution and cost-effectiveness, has emerged as a frequently used means for emotion recognition (ER). This paper presents a novel ensemble technique that comprises of 2 dimensional convolutional Neural Networks (2D-CNN) and a Transformer Encoder (TE). The proposed model first employs a 2D-CNN to efficiently extract location-specific and spectral properties directly from unprocessed EEG data, reducing the need for time–frequency transformations. Further, TE is used to model long-range dependencies by focusing on key segments of the input while enhancing computational efficiency through an optimized self-attention mechanism. The final classification layer predicts emotional states based on the extracted features. Evaluations on DEAP and SEED datasets demonstrate that the proposed ensemble paradigm achieves high accuracy, making it a robust and practical solution for BCI applications.