Due to the increasing number of amputees and the need to use prosthetics that simulate human limbs, an improved technique is proposed to classify hand gestures using Deep Recurrent Neural Networks (DRNN) based on the surface Electromyographic (sEMG) signal on the forearm. The implemented models are built on FeedForward Neural Networks (FFNN), Deep Recurrent Neural Networks (DRNN), and Long Short-Term Memory Networks (LSTM) using two types of datasets. They were recorded for four and seven motions, respectively. Both were written by MYO armband, and the conception of the technique is divided into two main phases applied to the two types of datasets. Two DRNN models are implemented, the First is a multi-classifications DRNN with all dataset files imported simultaneously. Each data file is then imported separately as input to the second binary classification DRNN model. Classification results for the multi-DRNN classifier and binary one is compared according to both datasets separately. Results show that the average accuracy for multi-classifications was (95%, and 86%) for both datasets while binary classification was 99% accurate for each model. Additionally, precision, recall, and f1-score were determined for both datasets, yielding better results.