For the always increasing amount of data new tools are needed to effectively harvest important information out of them. One of the core fields for data mining is Dictionary Learning, the search for a sparse representation of given data, which is widely used in signal processing and machine learning. In this paper we present a new algorithm in this field that is based on random projections of the data. In particular, we show that our proposition needs a lot less training samples and is a lot faster to achieve the same dictionary accuracy as state of the art algorithms, especially in the medium to high sparsity regions. As the spark, the minimum number of linear dependent columns of a matrix, plays an important role in the design of our contribution, we coined our contribution SparkDict.