Segmentation is one step of surface electromyography signal processing for pattern recognition. It consists of signal division in several windows, placed on the moments in which movements are performed. Then, these signals are sent for the feature extraction stage. However, the segmentation method and the feature extraction strongly interfere with the accuracy of the classifier.  In this premise, the objective of this work is to present a segmentation algorithm based on threshold analysis to identify the start and the end of the gestures automatically, which were named onset and offset. Parameters such as assertiveness rate, range of possible thresholds, and signal behavior were the variables inserted in the process of identifying the sEMG signal over time.  To evaluate the algorithm, a comparison with a similar study was performed, where the segmentation process was implemented using a traditional method (without being automatic) and applied to the recognition of six gestures. These gestures were: fingers flexed together in a fist, abduction of all fingers, tip pinch grasp, tripod pinch grasp, pointing index, and thumb up. Linear Discriminant Analysis (LDA) and k-Nearest Neighbours (KNN) classifiers were used with four sEMG features. The pattern recognition process achieved an average of 71.9% for LDA and 97.86% for KNN with the proposed segmentation method, being similar accuracies to similar work.