Word Spotting of Historical Arabic Documents is a challenging task due to the complexity of document layouts. This paper proposes a novel word spotting approach that consists of learning feature representation to describe word images. The objective is to investigate optimal embedding spaces to extract a discriminative word image representation. The proposed approach consists of two steps: i) construct a CNN-based embedding space with triplet-loss and then ii) match embedding representations using the Euclidean distance. For training, the CNN takes as input a set of triplet samples (anchor, positive sample and negative sample). Then, the triplet loss serves to create a novel space by minimizing intra-classes distances and maximizing inter-classes distances. The proposed approach is evaluated on the VML-HD dataset and the experiments show its effectiveness compared to the state of the art.