Electrocardiograms (ECGs) have shown unique patterns to distinguish between different subjects and present important advantages compared to other biometric traits, such as difficulty to counterfeit, liveness detection, and ubiquity. Also, with the success of Deep Learning technologies, ECG biometric recognition has received increasing interest in recent years. However, it is not easy to evaluate the improvements of novel ECG proposed methods, mainly due to the lack of public data and standard experimental protocols. In this study, we perform extensive analysis and comparison of different scenarios in ECG biometric recognition. Both verification and identification tasks are investigated, as well as single- and multi-session scenarios. Finally, we also perform single- and multi-lead ECG experiments, considering traditional scenarios using electrodes in the chest and limbs and current user-friendly wearable devices. In addition, we present ECGXtractor, a robust Deep Learning technology trained with an in-house large-scale database and able to operate successfully across various scenarios and multiple databases. We introduce our proposed feature extractor, trained with multiple sinus-rhythm heartbeats belonging to 55,967 subjects, and provide a general public benchmark evaluation with detailed experimental protocol. We evaluate the system performance over four different databases: i) our in-house database, ii) PTB, iii) ECG-ID, and iv) CYBHi. With the widely used PTB database, we achieve Equal Error Rates of 0.14% and 2.06% in verification, and accuracies of 100% and 96.46% in identification, respectively in single- and multi-session analysis. We release the source code, experimental protocol details, and pre-trained models in GitHub to advance in the field.