What is Transkribus?
Transkribus is a platform for (semi-)automatic HTR. Originally developed within the Horizon 2020 “READ” EU project, it is now administrated through the READ-COOP, who runs the servers and coordinates further development.
The usage for the platform is no longer fully free, but you will get 500 credits upon registration. This allows you to run text recognition on 500 handwritten (or 3000 printed) pages. If you want to continue, you will have to buy additional credits (at around 0.10 € per page).
All the image processing is done on the cloud servers of READ-COOP. Therefore, every scan has to be uploaded to the Transkribus infrastructure first.
To interact with Transkribus, there are two variants of the user interface (see https://readcoop.eu/transkribus/):
- Transkribus Lite which fully runs in the browser. It works best in Mozilla Firefox or Google Chrome and needs no installation. You can think of this application as the pendant to video-conferencing directly in the browser.
- Transkribus eXpert, a Java client that needs to be installed on your computer. This would be the pendant to installing a dedicated client for video-conferencing.
Functionality wise, the two interfaces are comparable, with a few more options in the eXpert client. Transkribus eXpert might also be a bit more stable and quicker to work with. Therefore, if you plan working with a larger number of pages, it might pay off to install and get acquainted with this client.
For a first introduction to Transkribus and for determining, how well a certain writing will be recognized, Transkribus Lite is an easy way to start. It is also helpful if you work from different computers and want to take a quick look at your documents.
Note: Transkribus is not the only option for Handwritten Text Recognition. eScriptorium (https://gitlab.com/scripta/escriptorium) and ocr4all (https://www.ocr4all.org/) are two other HTR-platforms under active development at academic institutions. But to our knowledge, there are no public instances yet, which would allow you to start working with them as easily as with Transkribus. In addition, all three solutions are based on comparable technologies. Getting acquainted with Transkribus should therefore prepare you for getting familiar with these alternatives very quickly.
Sign up
Independent of the client you plan to use, you need to create a user account for Transkribus first. This account provides access to the documents you plan to work with. You can use the same account for Transkribus Lite and eXpert. To get started, click on “Sign up for free” on https://lite.transkribus.eu/.
HTR process
Working with Transkribus is basically a three-step process:
- Upload the scans or photos either as a single image for every page or a single PDF file containing one or more pages.
With Transkribus eXpert, you can also import scans which are available on a public server directly through their URL without having them to download to your personal computer first. - In a second step, run the Layout Analysis for every page. This will recognize individual lines and – in the case of multi-column are multi-page scans – group the lines into areas such as a single column or page. For complex layouts with headings, footers, marginals or footnotes, as well in the case of insertions between the regular lines, manual adjustment of the layout might be needed.
- In a third step, you either manually transcribe every line or run the Text Analysis for every page. The quality of this step is highly dependent of the script and the specific hand of your source. Depending on these factors, one of the Public Models might lead to reasonable results. You will still have to do a manual post correction if you need a faithful transcription.
If the public models don’t work well enough – which means that more than 10-15% of the letters are wrongly recognized – you need to decide if you want to
- provide enough manually transcribed training material to train a custom model. This step usually requires a minimum of 20 to 30 faithfully transcribed pages. (https://readcoop.eu/transkribus/howto/how-to-train-a-handwritten-text-recognition-model-in-transkribus/)
1. Uploading Images / PDF
The basic entity of Transkribus are Documents, e.g. a letter or a diary. A Document consists of one or more Pages (Scans or Images might be a more precise designation in the case where we are working with images containing more than one page, for example in the case of a two-page scan of a diary).
You can group Documents into Collections. Use cases for a collection might be the correspondence between two persons with multiple letters exchanged, a personal diary which consists of multiple notebooks, or just all the documents related to a specific research project.
Collections can be shared with other users.

After uploading, you can set Metadata such as Title, Writer, Language:

You will find the link to this dialog in the Collection Overview after clicking on the three dots (…) in the lower right of a Document:

2. Layout Recognition
Before a page can be transcribed manually or by automated Text Recognition, all the lines containing text have to be recognized first.
You start this process from the Document Overview which shows a thumbnail view of every page. You can initiate this step either by selecting one or multiple pages individually or by clicking on Select all in the upper right corner:

Once you are done selecting at least one page, you can open the Layout Recognition Dialog by clicking on the corresponding Menu Entry in the TOOLS selection on the left:

Usually, the Presets in the Dialog (Baseline Detection) don’t need to be adjusted. Therefore, you can directly press [Start] to initiate this Job.

A confirmation dialog will be presented, which you confirm by again clicking [Start].
Depending on the number of pages you selected, Layout Recognition can be a rather quick operation. If you don’t see a result after 10 to 20 seconds per page, you can check the state of the Jobs in the TRANSKRIBUS ORGANIZER section in the left-sided menu:

Once you find the Layout-Analysis-Job as Finished, go back to the Document Overview.
Every page, for which Layout Analysis has been done, will then show an “Edit”-Button on the thumbnail:

But before actually start Editing the Text (or start the Text Recognition step), you first need to check the layout for correctness.
For this, you click on the respective page and then make sure to switch from the “View” into the “Layout” mode in the top-bar.

This will bring you to the page with purple indications of the recognized lines and text areas.

As you see here, reviewing the Layout depends on some global decisions, which should be thought through from the very beginning:
- How should we treat the Printed Letterhead (Cunard White Star, R. M. S. “Queen Mary”) at the top of the page? Should this be transcribed as well or do we leave it out?
Hint: It is always easier to start with more regions and remove them later, than deleting at this step and adding them back again in a later phase, after transcription has already been started.
You can select the areas to be removed by clicking in it and then press the Backspace or Delete-button on your keyboard (← Backspace / Del / entf / … depending on the language and/or brand).
In case you made a mistake, you can Undo by pressing [Ctrl]-Z (on a Mac: [command]-Z).
You can either select each area individually or select multiple areas at the same time holding [Ctrl] (on a Mac: [command]) while you click with the mouse-pointer on an additional area.
Once you are done, don’t forget to save your layout changes by clicking the green Floppy Drive in the upper left:

The second step is to check (and if needed correct) the order of the lines. To quickly check this order, it is helpful to display the reading order next to the areas.
You can activate this display by going to the Settings (lower left, gears, second from bottom) and the activate Show readingorder:

In case the lines are not numbered in the order you would expect, you have to open the Layout (bottom-most icon on the left), open the Region and drag the wrongly positioned line (e.g. Line 20) into the expected order (e.g. to position 19):

Another operation is merging multiple regions into one (when the analysis found multiple lines where there should be only a single one). For this, you select all the parts by at the same time holding [Ctrl] (on a Mac: [command]) while you click with the mouse-pointer on an additional area (in our Example Line 19 and Line 20).
Then right-click for the options-menu to show:

Select “Merge Shapes.” In a similar way, you could also split a single line into two parts (Horizontal / Vertical / Custom split).
Again, don’t forget to save your layout changes.
3. Text Transcription or Recognition
Once you have the layout of the page finalized, you can switch to the “Text” View in the top bar.
For every line defined in the layout, you now see a corresponding empty line, where you can start entering text.
1. Manual Transcription
You can start filling in the blank lines manually. Thanks to the Layout Recognition from the previous step, the corresponding line in the digital facsimile will be highlighted on the left side:

If a line is missing, you should go back to Layout and insert the line before continuing transcription beyond this step. Otherwise you risk a misalignment, which can have negative effects if you want to train a model later on.
2. Text Recognition
Instead of manual transcription, you can click on the split-T Icon between Layout and Date to get the Text Recognition Process started:

The crucial step is to select a suitable model. You should start with something that might correspond to the hand of your writer, which depends on the specific script (Kurrent / Sütterlin / Latin), period (early Modern or 20th century) as well as the cultural region (Dutch / English / French). Initially, you will have to browse through the list of public models and click on their Description to judge which Model might be a good fit:

In the end, only an actual test run can help you determine which model provides a reasonable result.
In our case, we could try:
- Transkribus German handwriting M1 (This model works well for documents/collections with Kurrent/Sütterlin and Latin script mixed. For only Kurrent or only Latin script the dedicated models might work better, so please test and compare before use.)
- or German_Kurrent_XIX-XX_M6-2 (Large train and test set for german kurrent (19. century).)
and then compare the results.
Once you selected a Model (indicated below the title of your Document), you can press [Start >]

In busy times of the day, the Text Recognition job might take up a couple minutes, if there are plenty of tasks of other users already lined up. You can again check the Jobs overview to find out if it has already ended. Once it shows as Finished, take a look at the page to check the results.
In our case, the model leads to quite good results, with a few errors that need manual adjustment.

If you find more than 5-10% of the characters misrecognized, you might want to try a different model (in this case e.g. German handwriting M1).
Even with a well matching model, a post-correction phase of the transcription will be needed.
In this step, you might also normalize certain visual appearances according to the transcription rules of your project. Examples might be quotes – if you differentiate opening and closing ones or unify them – and how you deal with certain letters such as the long s (ſ) – which could either be faithfully transcribed or normalized, as could certain ligatures (ff / fi).
In the case where you plan to train your model, you should stay faithful to the original wording, e.g. not silently expand abbreviations or normalize punctuation, since this will negatively affect the training process.
4. Provide enough manually corrected training material to train a custom model
If the public models don’t work well enough for transcribing larger collections of a writer, you can train a custom model. In this case, you need to come up with roughly twenty to thirty pages of clean transcriptions (the so called “ground truth”) of the same writer, which can then form the basis to re-train an existing model. After such a training step, you can expect a much lower error rate on the writer you trained the model for.
You can find an example of such a training process among the public models in Transkribus: Wolfthorn_Julie_final_base:KurrentM2:
Julie Wolfthorn’s handwriting based on letters to Ida and Richard Dehmel in the years between 1900-1936. Base model: German_Kurrent_XIX_M2

You can see how after a few training cycles; error rates quickly drop below 10 percent. In this case, 20’000 words have been manually transcribed or corrected. A smaller sample might already be enough for a significant improvement.
Additional Information
- https://readcoop.eu/transkribus/howto/getting-started-with-transkribus-lite/ is the “official” Getting Started Guide to Transkribus Lite.
- https://readcoop.eu/transkribus/resources/how-to-guides/ provides additional information on various topics such as Working with Tables or Training your own Models.
- The diary page displayed is taken from Gustav Geisel’s immigration diary from aboard the Queen Mary, 1938, https://collections.ushmm.org/search/catalog/irn517274#?rsc=139289&cv=0&c=0&m=0&s=0&xywh=-379%2C-125%2C2488%2C2488
OpenEdition suggests that you cite this post as follows:
migconn (November 11, 2022). Handwritten Text Recognition (HTR) in Transkribus Lite. Migrant Connections. Retrieved March 14, 2026 from https://migconn.hypotheses.org/140
1 comment