Early arrhythmia detection helps in a timely treatment of patients to increase their life expectancy. Arrhythmia detection is performed using standard 12-lead electrocardiogram (ECG) signals by cardiologists. In this work, we used a fifteen layer Residual Neural Network (ResNet) that takes a standard 12-lead ECG segment and classifies between eight different types of concurrent arrhythmias and normal sinus rhythm. The ECG segments were divided in equal duration of six seconds each as the neural network ingests a constant sized input and the minimum duration of an ECG segment present in the dataset was of six seconds. The segments of length 3000 samples obtained from the recording of the patient are labeled corresponding to the recording. The employed ResNet model encompasses a variety of layers including Batch Normalization, Dropout, 1-Dimensional (1-D) Convolution, Max Pooling, and Fully connected layers. Batch Normalization and Dropout are used to reduce overfitting and improve generalisation. 1-D Convolution takes into account the temporal structure present in the ECG signal and the residual connections allow the gradient to backpropagate to the initial layers so that the initial layers are able to mitigate the problem of vanishing and exploding gradient. Max Pooling, and Fully connected layers are also used to increase the learning capacity of the model. We used batch wise training as they are computationally more efficient than traditional stochastic gradient descent. As the size of training data is huge, batches provide us the liberty to not have all the training data in memory. The trained ResNet model provided F-beta and G-beta scores of 0.628 and 0.426, respectively on the test data.