Deep learning has advanced electromyography (EMG) based gesture recognition, yet existing models face robustness challenges. We propose a novel Transformer-based framework to enhance classification accuracy. Our architecture introduces two key innovations: a customized Patch Embedding module to adapt 1D time-series EMG signals for self-attention, and a novel Parametric Tanh Activation Function (DyT) that replaces conventional Layer Normalization to improve training stability and generalization. We evaluated our model on two public datasets representing distinct modalities: a sparse sEMG dataset (NinaPro DB2 Exercise B) and a high-density EMG dataset. The framework achieved high classification average test accuracies of 85.18% and 86.69%, respectively. These results confirm the effectiveness of our architecture, demonstrating its significant potential for real-world applications such as prosthetic control and human-computer interaction.