The Muscle Computer Interface (MCI) system is one of the areas of active interest in virtual reality and human-computer interaction research. The main problem associated with the MCI was the EMG signal classification, to facilitate the effective combination of an MCI system with deep learning methods. Surface EMG signals include high-density transient signals and sparse multi-channel signals. The former was analogous to an image that can be recognized by a CNN network. The latter was studied in this investigationin which an MCI system with an MYO armband was realized. Sparse multi-channel EMG signals were long-term sequence signals with a high correlation between time and time that can be recognized by an RNN network. We proposd a combined RNN network architecture to recognize gestures with multi-stream feature sequence signals that were obtained by extending the original signals in the time-domain and time-frequency domain. The accuracy of the net is 90.78%. We perform cross-validation without a self-training set using 35 individuals, and the accuracy of the classification is 78.01%. The accuracy of real-time gesture recognition in the MCI system is 82.09%, and the action can be recognized within 61.7 milliseconds. We establish that the combined RNN nets can classify different gestures using EMG signals, and the MCI system performs well in generalization and real-time recognition.