Millions of people worldwide are affected by arrhythmias. Arrhythmias are abnormal activity of the heart functioning. Some arrhythmias are harmful to the heart and can cause sudden mortality. The electrocardiogram (ECG) is a significant tool in cardiology for the diagnosis of arrhythmia beats. Computer-aided diagnosis (CAD) systems have been proposed in several studies to automatically classify different types of arrhythmias from ECG signals. To improve the classification of arrhythmias, a new end-to-end feature learning and classification model has been developed. This work focuses on the implementation of a one-dimensional convolution neural network (1D-CNN) model based on an auto-encoder convolution network (ACN) that learned the best ECG features from each heartbeat window. After that, we applied a Support Vector Machine (SVM) classifier for auto-encode features in order to detect the four different types of arrhythmic beats, including normal beats. These arrhythmia beats are left bundle branch block (L), right bundle branch block (R), paced beats (P), and premature ventricular contractions (V). using the MIT-BIH arrhythmia database. The statistical performance of the model is evaluated using tenfold cross-validation strategies and obtained as an overall accuracy of 98.84%, average accuracy of 99.53%, sensitivity of 98.24% and precision of 97.58%, respectively. This model has presents better results than other state-of-the-art models. Therefore, this approach may also help in clinical heart care systems.