Gesture recognition system can be an important tool to assisting hearing impaired people, as games, mobile applications, virtual reality, and others. This work presents a classification approach considering Attention Mechanism (AM) in Recurrent Neural Networks (RNNs) for Sign Language Recognition based on data gloves. The objective of this work is to present a comparison between RNNs with and without AM. Three gloves database, composed of flex and inertial sensors, are utilized in this work. The first database consists of 10 Brazilian Sign Language (Libras) words; the second database is composed of 26 Brazilian Sign Language (Libras) alphabet; and third comprises of 16 ASL words. The RNNs considered were Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). All the classifications were employed with leave-one-subject-out (LOSO) approach, and the mean accuracy for the volunteers are presented as final result. All the models were tested with different number of hidden units. For all analyzed database, the insertion of AM in the RNNs improved the final accuracy. The statistical analysis showed that, for some volunteers, the results with AM are significantly superior.