Compressed sensing (CS) is one of the predominant tools used presently to explore the possibilities in accelerating cardiac and body Magnetic Resonance (MR) imaging for achieving shorter scans and accommodating a wider patient group. CS accomplishes this by manoeuvring the scanning hardware to make much fewer measurements and imposing sparsity of the MR image in a known basis on the reconstruction process. Prior works have adopted fixed transforms as well as dictionaries learned on image patches to incorporate the sparsifying basis. Despite the obvious merits of 1-D dictionary learning methods, they suffer from high computational and memory complexity when extended to higher patch sizes and are thus restricted to capturing only local features. Thus, there arises the need for an efficient framework to extend the advantages of dictionary learning to higher dimensional applications like dynamic MRI (dMRI) where there exists a strong correlation across successive frames. This work employs a tensor decomposition based dictionary learning approach to effectively extend CS to dMRI and exploit the temporal gradient (TG) sparsity, thereby retaining maximum spectral and temporal resolution at higher under-sampling factors. The proposed technique is experimentally validated to achieve significant improvement in reconstruction quality over the current state-of-art in dynamic MRI under distinct sampling trajectories and noisy conditions. Further, it facilitates faster reconstructions of the dMRI volume than existing methods, rendering it an ideal choice in critical scenarios which demand a swift diagnosis.