Segmenting documents into lines and words is a very critical step before the recognition task. It is even more difficult for ancient and calligraphic writings, as is often the case in Arabic manuscript documents. In this work, we propose a new attempt to segment documents into lines and words, using deep learning. For line segmentation, we use an RU-net which allows a pixel-wise classification, thus separating pixels of lines from the background pixels. For segmenting lines into words, not having a ground truth for the word segmentation (at the image level), we use the line transcription to find the words. A BLSTM-CTC is used to achieve this mapping directly between the transcription and line image, without any segmentation. A CNN precedes this sequence to extract the features and feeds the BLSTM with the essential of the line image. Tested on KHATT Arabic database, the system achieves good performance that is of the order of 96.7\% correct lines and 80.1\% correct words.