Teach Transkribus a ledger’s layout before exporting its handwriting

A register can give one object three lines of description while keeping its date on a single line. In the spreadsheet, those lines still belong to one record. Transkribus addresses that structure with a trainable table model: separate models detect row and column boundaries, then line detection and handwriting recognition work inside the resulting cells.[1]

This workflow is for a volume with a recurring grid, such as a handwritten loan register. It follows the current product documentation and ends with an inspected spreadsheet export. The register example below is invented to explain the check; no recognition run is claimed.

Choose the pages and check access

Use the Transkribus Web App with access to Table Recognition and advanced spreadsheet export. Scholar, Team and Organisation are the paid routes identified in the current documentation. Processing also consumes credits; check the estimate before each job because table recognition, line detection and text recognition are distinct operations.[2][9]

For a simple recurring layout, start with around 20 annotated pages; complex layouts need more. Choose examples across the volume so different row heights, handwriting and page conditions are represented.[1][4]

Open the collection that should hold the register and choose Upload. Add its JPEG, PNG or TIFF images, or a PDF; give the document a title and select Submit. A PDF becomes a sequence of page images. Keep filenames and page order meaningful so you can find a source entry again after export.[3]

Teach the recurring grid

For the initial training pages, use one table per page and no merged cells, as the table-model instructions require. Ground Truth means the corrected examples that you supply for learning.[4]

1. In the editor, choose Add Region > Add Table Region. Click to mark the beginning and end of the table.
2. Select the table. Hold V and click the column boundaries; hold H and click the row boundaries.
3. Check every cell, then save the page as Ground Truth. Repeat for the chosen training pages.
4. Select the prepared documents in the collection. Open the dropdown and choose Train Model > Table Model; restrict the selected data to the intended Ground Truth pages.
5. Give the model a descriptive name and target collection. Retain the recommended initial defaults, review the setup and choose Start training.

Keep a few representative pages aside for your first practical check. A two-line description, a blank cell and a change in the writer’s spacing are more revealing than another page with only short, evenly placed entries.

Run three jobs in the right order

Select the check pages and choose Process with AI > Table Recognition. Choose your model, review the credit estimate and start recognition. Inspect the grid before continuing.[4]

Next run layout recognition separately. Under Process with AI, choose the layout process—called Layout Analysis in the general layout instructions—and open Advanced Settings. Select Keep existing Text Regions and Split lines on region border. The second option appears when the first is selected. These settings preserve the table regions and prevent nearby writing in adjacent cells from becoming one continuous line.[5]

A baseline is the line along the bottom of a handwritten text line. After layout recognition, inspect those lines on the image: a long description can have several baselines within one cell. It should not acquire an extra table row just because its writer continued on the next line.[6]

Then return to Process with AI and choose Text Recognition. Select a model appropriate to the register’s script, language and period, review the credit estimate and start the job. Open the processed pages in the editor to see the transcript beside the images.[7]

Correct the text and its cell placement before exporting. The product’s table walkthrough puts this editing stage after recognition. If the grid is wrong across many pages, improve the training examples and model before committing the rest of the volume to the same error.[8][1]

Check a record that spans several written lines

Imagine one row in our fictional loan register: the date is “7 May”; the object description reads “Map of East Wharf”, written on two lines; and the return-date cell is empty. The export should contain one record with the whole description and an empty return date.

A separate record containing only “East Wharf” reveals a structural split, even if every letter was recognised correctly. A return date borrowed from the following row is another structural error. Check the page’s cell boundaries and baselines first, then its transcription. Correcting letters alone will not fix membership in the wrong record.

Export the checked pages

Select those pages, open Action > Export and choose Spreadsheet (Excel/CSV), then Table Export. Start export. Merge into one table can combine selected tables across documents; use it only when the columns have the same meaning and order. The export options also include cell-image snippets for checking.[9]

Find the finished export in Processes & Activity under Up- & Downloads, open its three-dot menu and choose Download. Open the saved file in your spreadsheet application. Transkribus also emails a download link, valid for two weeks, so retain a local copy you can return to.[9]

Compare the first and last records, a multi-line description and every deliberately blank cell in your small check set. Confirm that records stay together when you move down the sheet. When a correction is needed, make it in the Transkribus source and export again, keeping the page image beside the decision. Once that sample behaves as intended, process the remaining matching pages with the same sequence.

Report an error or suggest a correction