Atrial fibrillation (AF) and atrial flutter (AFL) represent atrial arrhythmias closely related to increasing risk for embolic stroke, and therefore being in the focus of cardiologists. While the reported methods for AF detection exhibit high performances, little attention has been given to distinguishing these two arrhythmias. In this study, we propose a deep neural network architecture, which combines convolutional and recurrent neural networks, for extracting features from sequence of RR intervals. The learned features were used to classify a long term ECG signals as AF, AFL or sinus rhythm (SR). A 10-fold cross-validation strategy was used for choosing an architecture design and tuning model hyperparameters. Accuracy of 88.28 %, with the sensitivities of 93.83%, 83.60% and 83.83% for SR, AF and AFL, respectively, was achieved. After choosing optimal network structure, the model was trained on the entire training set and finally evaluated on the blindfold test set which resulted in 89.67% accuracy, and 97.20%, 94.20%, and 77.78% sensitivity for SR, AF and AFL, respectively. Promising performances of the proposed model encourage continuing development of highly specific AF and AFL detection procedure based on deep learning. Distinction between these two arrhythmias can make therapy more efficient and decrease the recovery time to normal heart rhythm.