This work investigates accelerating CEST imaging using patch-based global spatial-temporal dictionary learning (G-KSVD). We extend the dictionary learning for CEST acceleration. CEST data has high spatial-temporal correlation, so we can utilize the global Z-Spectrum information as well as the spatial information to form the global spatial-temporal dictionary. The dictionary is learned iteratively from overlapping patches of the dynamic image sequence along both the spatial and temporal directions. The proposed method performs better than the BCS and k-t FOCUSS methods for both phantom and in vivo brain data at high reduction factor of R=8.