In this work, we present a multi-scale analysis dictionary learning framework in the presence of weak supervision. The motivation for using multiple scales on the dictionary atoms is that data has patterns that are best captured at different scales in many cases. As such, uniscale dictionary learning will have difficulty in capturing different patterns of interest, especially for convolutive modeling of the data. We propose a probabilistic graphical model with a multi-scale dictionary and develop an inference framework for the proposed model. To evaluate our proposed approach, we conduct experiments on both synthetic and real-world data. The results show that the proposed multi-scale dictionary learning outperforms the uni-scale approach when data contains multi-scale patterns, and the performances are comparable when the data is uni-scale.