This study proposes DeepWriteSYN, a novel on-line handwriting synthesis\napproach via deep short-term representations. It comprises two modules: i) an\noptional and interchangeable temporal segmentation, which divides the\nhandwriting into short-time segments consisting of individual or multiple\nconcatenated strokes; and ii) the on-line synthesis of those short-time\nhandwriting segments, which is based on a sequence-to-sequence Variational\nAutoencoder (VAE). The main advantages of the proposed approach are that the\nsynthesis is carried out in short-time segments (that can run from a character\nfraction to full characters) and that the VAE can be trained on a configurable\nhandwriting dataset. These two properties give a lot of flexibility to our\nsynthesiser, e.g., as shown in our experiments, DeepWriteSYN can generate\nrealistic handwriting variations of a given handwritten structure corresponding\nto the natural variation within a given population or a given subject. These\ntwo cases are developed experimentally for individual digits and handwriting\nsignatures, respectively, achieving in both cases remarkable results.\n Also, we provide experimental results for the task of on-line signature\nverification showing the high potential of DeepWriteSYN to improve\nsignificantly one-shot learning scenarios. To the best of our knowledge, this\nis the first synthesis approach capable of generating realistic on-line\nhandwriting in the short term (including handwritten signatures) via deep\nlearning. This can be very useful as a module toward long-term realistic\nhandwriting generation either completely synthetic or as natural variation of\ngiven handwriting samples.\n