Sign languages are the main form of communication for the hearing-impaired population. Still, due to the lack of knowledge on the part of most of society about the gestures used, this population has limited integration. In this scenario, sign language recognition systems can represent a viable way of expanding this form of communication to everyone. With this in mind, this work compares different time series recognition techniques, aiming at an analysis of Deep Learning (DL) techniques in surface electromyography without feature extraction. In this work, classification models based on the Time Series Perceiver (TSPerceiver) and Gated Multilayer Perceptron (gMLP) techniques were trained over separated subject databases with average accuracies between classes up to 88.4% and 94.6%, respectively, in addition to comparisons with these same models trained over the entire database, which achieved accuracies of 72.4% and 60.8% respectively.