<p>The self-executing investigation of electrocardiogram (ECG) signal plays a pivotal role in the early identification and supervision of cardiac arrhythmias. The diversified physiology of arrhythmia and the presence of precise dissimilarities in the chronic ECG characteristics create obstacles in conspiring authentic automated methods. The elimination of artefacts, peak dislocation, and unattended weaker neighbouring packets with strong spectral neighbours usually pose a challenge during the process of arrhythmia detection. To deal with these challenges, this paper exploits the Fractional Superlet spectral analysis (FSSA) that smooth out the variation in time–frequency resolution as a function of frequency by allowing fractional order of wavelets and hence makes the signal noise-free and ensures proper placement of peaks. The clear depiction of time–frequency pattern enhances the accuracy of heart rate estimation. It also presents a deep learning-based model employing a time–frequency energy representation of the signal to detect and classify multiple classes of arrhythmia. This provides enough information to make preferable use of CNN performance thus eliminating the need for artefact removal and feature extraction steps. The proposed method is validated over MIT-BIH Arrhythmia DB, MIT-BIH Malignant Ventricular Ectopy DB, MIT-BIH Atrial fibrillation DB, and MIT-BIH Supraventricular DB, to scrutinize the performance. An ensemble of the FASLT technique and VGG19 yields an overall accuracy, sensitivity, specificity, precision and F1-score of 98, 97.95, 99.32, 97.96 and 97.93% respectively for arrhythmia detection. The proposed method performs better in terms of specificity in comparison to existing methods.</p>