Deep neural networks has shown great success incomputer vision fields by achieving considerable state-of-the-artresults and are beginning to arouse big interest in the documentanalysis community. In this paper, we present a novel siamesedeep network of three inputs that allows retrieving the mostsimilar words to a given query. The proposed system followsa query-by-example approach according to a segmentation-based technique and aims to learn suitable representations ofhandwritten word images, for which a simple Euclidean distancecould perform the matching. The results obtained for the GeorgeWashington dataset show the potential and the effectiveness ofthe proposed keyword spotting system