This paper delves into the significance of employing neural networks (NN) for the classification of sEMG signals, given their inherent complexity. The methodology entails recording muscle signals from 10 volunteers performing three distinct movements. Subsequently, 9 of these records were utilized for training the NN, and 1 was reserved for evaluating the efficacy of each of them (leaving one subject out). Two approaches were employed in constructing the NN: Long Short-Term Memory (LSTM) and Multi-Head Attention (MHA). In the first model, the input data consisted of feature extractions obtained through Fast Fourier Transform (FFT), whereas in the second approach the input data were directly the FFT signal. A Data Augmentation (DA) method was implemented to expand the available dataset and to ensure the gesture classification efficiency of each model. The results obtained demonstrated that the usage of data augmentation was efficient in generating new data for the training of both NN. The LSTM approach exhibited superior results when trained with data augmentation. However, when compared to the results obtained by leaving one subject out, the impact of data augmentation on the final results was not significant. However, the MHA approach achieved the best results in leaving one subject out of validation when trained without data augmentation, but having the best results during the training of the models. The results obtained after the statistical test of Wilcoxcon indicate that the usage of DA does not generate side effects since the accuracy distribution for the models trained with DA and without DA was the same.