Environmental Sound Classification (ESC) is a challenging task due to the diverse and highly variable nature of real-world acoustic environments. This paper proposes a Two-Path Deep Attention Network based on a Transformer Encoder to enhance feature representation and classification robustness. The proposed architecture employs two parallel attention branches to independently learn complementary acoustic features before feature fusion, enabling more comprehensive representation learning. In addition, Speed Perturbation and Vocal Tract Length Perturbation (VTLP) are applied as data augmentation techniques to improve model generalization. Experiments conducted on the ESC-50 and ESC-10 benchmark datasets demonstrate the effectiveness of the proposed method, achieving classification accuracies of 93.35% and 97.15%, respectively. The results show that the proposed two-path architecture combined with full data augmentation consistently outperforms conventional single-path CNN-, RNN-, and Transformer-based approaches, confirming its robustness for environmental sound classification under complex acoustic conditions.