Keyword spotting (KWS) in historical documents is an important tool for the initial exploration of digitized collections. Nowadays, the most efficient KWS methods rely on machine learning techniques, which typically require a large amount of annotated training data. However, in the case of historical manuscripts, there is a lack of annotated corpora for training. To handle the data scarcity issue, we investigate the merits of self-supervised learning to extract useful representations of the input data without relying on human annotations and then use these representations in the downstream task. We propose ST-KeyS, a masked auto-encoder model based on vision transformers where the pretraining stage is based on the mask-and-predict paradigm without the need for labeled data. In the fine-tuning stage, the pre-trained encoder is integrated into a fine-tuned Siamese neural network model to improve feature embedding from the input images. We further improve the image representation using pyramidal histogram of characters (PHOC) embedding to create and exploit an intermediate representation of images based on text attributes. The proposed approach outperforms state-of-the-art methods trained on the same datasets in an exhaustive experimental evaluation of five widely used benchmark datasets (Botany, Alvermann Konzilsprotokolle, George Washington, Esposalles, and RIMES).