Abstarct: Electrocardiogram (ECG) classification is a vital task in the diagnosis and treatment of cardiac abnormalities. ECG signals provide valuable information about the state of the heart, and the classification of these signals as normal or abnormal plays an important role in the diagnosis and treatment of various cardiac abnormalities. In recent years, various feature extraction methods for automatic ECG classification have been developed in order to increase the accuracy and efficiency of ECG classification. These methods use different signal processing techniques. In this thesis, a robust feature extraction method from ECG signal using normalized Capstral coefficients is proposed to detect cardiac arrhythmias. The features extracted from the proposed method are given to a CNN network for classifying signals. Also, different parameters of the CNN network have been investigated to select the optimal network. With the help of statistical study, it has been shown that the proposed feature extraction method and the presented network have a suitable and high accuracy in detecting unhealthy signals of ventricular arrhythmia and malignant ventricular ectopy from healthy signals. The results of the experiments show that the network has 98 percent accuracy in the test phase.