In the field of image processing, document-level image processing is one of the trending topics. Various researchers have proposed works on Bangla handwritten documents, and one of them is segmentation and ground truth generation. Segmentation schemes for lines, words, and characters are a crucial preprocessing task in many applications. Segmentation algorithms for Bangla handwritten offline documents that have been proposed so far were unable to achieve significantly high accuracy due to various problems like overlapping of components, cut-through, etc. Fully manual segmentation can produce highly accurate results but it is not feasible for any user. In this paper, a semi-automated system has been proposed for Bangla handwritten documents which can potentially produce fully accurate segmentation results. Furthermore, ground truth generation for words has also been added to the system using the isolated character recognition CNN model.