Although most people can communicate effectively through speech, some have difficulties doing so due to physical or mental impairments. Communication is a significant obstacle for individuals with these disabilities. Methods of deep learning can aid in the elimination of communication barriers. This article proposes a model based on deep learning for detecting and recognizing words from gestures. Deep learning models such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) are used to recognize signs from Egyptian Sign Language (ESL) video frames. There are many activation functions, and every type has advantages and disadvantages. All these activation functions were applied to our dataset for ESL. To overcome the main disadvantage of the Relu activation function, we proposed a gesture recognition method for ESL using Mediapipe and modified GRU with a new activation function