In this paper we present an application of a recurrent neural network (RNN) to recognize gestures of Microsoft Research Cambridge 12 (MSRC-12) dataset. Three different processing methods of joint coordinates are used in the artificial neural network, our objective is to specify which method results in a more accurate network. The MSRC-12 dataset is captured by the Kinect sensor, it consists of a sequence of human body articulation movements. In addition, the FastDTW algorithm is employed to normalize the data frames number. The three different methods proposed in this paper are: the 3D coordinates method, the subtraction method, and the normalization method. These three methods are used in a RNN model, and we obtained with 3D coordinates method an accuracy rate of 87,30%, using subtraction method the ac-curacy rate is 87,11% and with the normalization method the accuracy rate obtained is 89,14%.