Electrocardiogram (ECG) beat classification plays an important role in diagnosing and managing cardiovascular conditions. This study introduces a method for classifying ECG beats into normal, supraventricular, and ventricular types using a transformer-based architecture with a custom attention mechanism. The preprocessing steps include filtering, resampling, normalization, and segmentation of ECG signals. Two classification approaches are proposed: (a) a cascade method employing sequential binary classifications, and (b) a multiclass method for direct classification. Synthetic data simulating supraventricular and ventricular arrhythmia is generated to address data imbalance and enhance model performance. The training and testing processes leveraged various datasets, including the MIT-BIH Arrhythmia, Supraventricular Arrhythmia, European ST, and Long-Term ST databases. The results highlight the superior performance of the cascade model, achieving a macro-average F1 score of 0.92 and a precision score of 0.90 on the Long-Term ST database. These findings show the importance of dataset selection and enhancement, demonstrating the model’s potential for clinical applications.