Objective: The diagnosis of cardiovascular diseases is a lengthy and expensive procedure. The automated detection of cardiac abnormalities can increase the chances of early prevention and successful treatments. In this study, we present a methodology for multi class arrhythmia detection and classification over arbitrary length 12-lead electrocardiogram (ECG) recordings. Approach: To this end, we proposed a 1-D convolutional neural network (CNN) architecture, which learns patterns directly from raw 12-leads ECG signals to classify different rhythms, while eliminating the necessity to detect beat locations and to extract the traditionally used hand-crafted features. At first baseline wander removal and Butterworth filter for each channel in ECG are applied as a preprocessing stage. ECG segments of 30 seconds are then fed to a 1D CNN by adding noise in the training data to improve the robustness of the model. Once the 1D CNN is trained, it can be used to classify 12-leads ECG data segments in a fast and accurate manner. Such a solution can also be used for real-time ECG monitoring on a light-weight wearable device. Results: The proposed methodology is evaluated using the scoring mechanism provided by PhysioNet/Computing in Cardiology (CinC) Challenge 2020. We obtained an F_beta score of 76.2±1.7% and a G_beta score of 55.4±2.2% over the training dataset with 5-fold cross-validation. On the unseen test dataset our algorithm has achieved the corresponding scores of 77.1% and 56.7% respectively. Conclusion: The proposed approach initially showed a promising performance in distinguishing between the different rhythms without the need of any hand-crafted features. This opens the horizon for computer automated interpretation of 12-leads mobile ECG.