Transkribus Handwriting Recognition for Old Records

Transkribus handwriting recognition is a web platform that applies a trained text model to images of handwritten or printed pages and returns a transcript you correct beside the image. Its help center describes it as made for historical documents. For old legal records it can supply a first draft, which a person must still check.

Everything said about Transkribus in this post comes from its own help center, opened and read for the purpose. The post does not rate how well any model reads any document. Transkribus itself says results depend on matching the model to the material, so the only useful evidence is a trial on your own pages.

What is Transkribus handwriting recognition?

Transkribus is an online service for transcribing documents, run by a cooperative called READ-COOP. Its help page on what you can do with Transkribus lists transcribing "historical and modern handwritten and printed text", editing layout and text, training models on a specific handwriting style, collaborating, and exporting.

The working unit is a text model. The help page on automatically transcribing your documents defines a text model as an algorithm trained on images and their transcriptions, and says its purpose is to determine the most likely sequence of characters for each section of handwritten text. You upload pages, pick a model, start a recognition job and open the result in an editor.

The steps the help center gives

1. Upload images or PDFs to a collection. The uploading page accepts JPEG, PNG and TIFF images and PDF files, and says a resolution of around 300 dpi is sufficient.
2. Select the pages and click "Process with AI".
3. Choose a text model that matches the language and period.
4. Check the number of credits the job will use, which is displayed before you start.
5. Click "Start recognition", then open the page when the job has finished.

The same page recommends that first-time users test recognition on a few pages before processing a whole document. That is sound advice for a firm with a box of old minute books: pick three pages that represent the best, the typical and the worst, and judge from those.

During recognition the page is divided into text regions and lines. How software finds lines and reads them is the subject of the post on handwritten text recognition.

Choosing a Transkribus text model by language, script and period

You choose a Transkribus model by filtering for the language, the period and whether the material is handwritten or printed. The help page on choosing a model says plainly that no general model exists for all handwriting, and that specialised models are expected to remain necessary.

That page lists what to weigh when you pick a text model:

• the type of material, handwritten or printed
• the language
• the period
• the type of script
• the Character Error Rate shown for the model

Public models

Public models are ready to use. According to the help page on public models, some were made by the READ-COOP team and most were trained by Transkribus users who chose to share them. You can filter the list by century, language, material and script, open a model's description, and upload a test image to try it before you spend anything on a full job.

The help center is candid about the limits. Its FAQ on finding the best public model for your documents says public models "will not always yield a perfectly recognised text, as handwritings and material can differ."

Reading the Character Error Rate

Each model's page reports a Character Error Rate, or CER. Transkribus's page on the Character Error Rate and learning curve defines it as the percentage of characters the model transcribed incorrectly, and gives the example that a 5% CER means 95 characters in 100 were right. It also says the published figure is measured on the model's validation pages, which are transcribed pages set aside from training, and that a model applied to hands it did not see in training "may perform worse".

For a lawyer, that last point is the one to hold on to. The CER on a model's page describes the documents the model was built from. It does not describe your county's deed books. And a small percentage is still a lot of characters on a full page, some of which will fall in names, dates and sums.

Private models

If no public model suits, you can train your own. The page on data preparation says to transcribe at least 25 pages first, and advises at least 10,000 words for each hand in handwritten material. These accurate transcriptions are called Ground Truth, and the page warns that any mistake in them teaches the model something wrong.

That is real work. For a single volume you need to read once, typing the pages you need may cost less than training a model. For a long series in one clerk's hand, such as decades of a court's order books, a trained model can repay the effort.

Super Models

Transkribus also offers what it calls Super Models. Its page describes them as transformer-based models with an optical part that processes the image and a language model that tries to make sense of and improve the extracted text, able to handle handwritten and printed text in one pass. The page says users cannot currently fine-tune them, and that a specialised model trained for well-defined material can still yield better results.

How do Transkribus credits work for handwritten pages?

Transkribus charges recognition in credits, and a handwritten page uses more than a printed one. Its credit usage overview lists handwritten text with lines at 1 credit and printed text with lines at 0.5 credits. Uploading, downloading, manual transcription and model training are stated to be free.

The overview page breaks usage down by the type of recognition as well:

The page adds that table or field recognition also needs line recognition and text recognition for a complete transcription, so an estimate for a ledger should count all three.

Both that page and the page on managing credits say a new account receives 50 free credits a month for text recognition, so that you can explore and test. Monthly credits expire at the end of the subscription cycle, according to the managing credits page, while add-on credit packages do not expire. For what credits cost, read the Transkribus pricing page. This post states no prices because they are the vendor's to change.

Estimating a job

The recognition window shows the credits a job needs before you start it, so you never have to guess. A rough plan for a firm:

1. Count the pages you actually need. A 300-page minute book may hold 12 pages that concern your easement.
2. Run those pages, not the volume.
3. Use the free monthly credits for the trial pages, and compare models on the same page before you run the rest.

Because the free allowance is monthly and counted in credits, a small research task may fit inside it. Check the current allowance on the vendor's pages before you count on that.

Correcting a Transkribus transcript in the editor

You correct a Transkribus transcript in its Document Editor, with the page image on one side and the text on the other. The help page on transcribing manually says clicking a line in the image highlights the corresponding line of text, so you check line against line.

Features of the editor that matter when the document is evidence of title or of an old judgment:

• Line-by-line alignment. The image centres on the line you are working on, which keeps your place in a dense page.
• Reading order. In layout mode you can drag lines or text regions into the correct order when the software has read a marginal note in the wrong place.
• Comments. You can add comments in the text editor's sidebar.
• Version history. Each save is stored as a new version of the page. You can restore a version, and you can compare two versions with the differences highlighted in colour.

Version comparison is useful in two ways. Run two models on the same page, and the comparison shows where they disagree. Those are the words to look at hardest. Later, it shows what a reviewer changed in a colleague's transcript.

One point from the training pages is easy to miss. The help page on training text recognition models says that simply correcting recognised text in the editor does not improve the model. Your corrections fix that page. The model reads the next page exactly as it would have before, unless you save corrected pages as training data and run a new training job.

What to check first

Start with the words where a wrong reading does the most harm in an old record: surnames, place names, dates, sums of money and abbreviations. Read those against the image before anything else. If the hand itself is unfamiliar, the post on how to read old handwriting sets out a method and points to free archive guides.

Which formats can a Transkribus transcript be exported in?

Transkribus exports transcripts as Word, PDF, plain text and several XML formats. Its downloading page lists the standard options available on all subscription plans, and more configurable options on the Scholar, Team and Organisation plans.

The options the page lists for the paid plans include spreadsheet export of tables and tags, ALTO XML, TEI XML and METS, along with settings for the Word and PDF exports.

Settings worth knowing for legal use

• Mark unclear words. If words in the transcript have been tagged "unclear", the Word export can highlight or mark them. Tag doubtful readings as you correct, and the reader of the Word file sees where the transcriber was unsure.
• Preserve line breaks. The Word file keeps the original line breaks, which makes it easier to compare with the page.
• Force page breaks. Each page of the document starts a new page in Word, so page references stay aligned.
• Image plus text layer. The PDF holds the page images with a searchable text layer. The page warns that the highlighted area in a search may not perfectly match the word in the image, because word positions are worked out from the lines.
• Extra text pages. The PDF gains a page of transcribed text after each original page.
• PDF type. Standard PDF or PDF/A.

Export is a server job. The downloading page says you receive an email with a download link that is valid for two weeks. Plan for that if the transcript is needed by a deadline.

Is Transkribus a fit for a law firm's historical records?

Transkribus fits a firm whose records are old, handwritten and not confidential in themselves, and that has someone able to check the result. For a current client's handwritten papers, the questions are about confidentiality before script. The vendor's statements help you decide, and your own review still comes first.

Where it suits

• Recorded deeds, probate registers, court minute books and parish or municipal records that are already public.
• A long run in one or two hands, where a matching public model exists or training a private one is worth it.
• Work where a searchable draft is the goal, such as finding every mention of a surname across a volume before you read those pages properly.

For the particular demands of one conveyance, see the post on how to transcribe a handwritten deed.

What Transkribus says about data

The FAQ titled Is my data safe? states that you retain ownership of your documents and transcriptions and can delete them at any time, and that "All data is processed and stored on our own servers in Austria." The uploading page says documents are private by default and stored on READ-COOP's servers in Innsbruck. The FAQ also says: "By using the platform, you give us permission to process your data so that we can provide the service and improve it."

Those are the vendor's words, quoted so you can weigh them. Before client material goes in, a firm should ask:

• What does permission to "improve" the service cover, and does it include training on uploaded pages?
• If you train a private model on a client's documents, who can use that model, and can it be deleted?
• Who in your collection can see each document, and how is access removed when a matter closes?
• What do the terms and conditions the FAQ links to say about each of these?

The post on whether it is safe to upload client documents to an OCR tool gives a fuller list of questions for any service.

A worked example

Okafor and Lindqvist represents a landowner in a right-of-way dispute. The chain of title runs through an 1887 deed and a set of commissioners' minutes from the 1890s, both held as images by the county archive and open to the public. An associate:

1. Uploads six pages: the deed and the five minute-book pages the index points to.
2. Tries two public models filtered for English, nineteenth century, handwritten, on one page each, and compares the versions.
3. Runs the better model on all six pages using the free monthly credits.
4. Corrects each line against the image, tagging four words as unclear.
5. Exports Word with line breaks preserved and unclear words marked, and a PDF with a text layer.
6. Has a second person read the names, dates and the description of the way against the images.

The draft saved typing. The checking in steps 4 and 6 is what makes the transcript usable.

On the deed and the minute book, check the parties' names, the dates, the amounts, and any signature the clerk copied. You are following a chain of title into an easement question. Give the distances the same care, because a quiet title claim or an encroachment allegation will use those lines. In a boundary dispute, compare every number with the image.

When the records are not historical

A modern typed production with some handwriting on it is a different job. A general OCR service reads the typed pages, and you deal with the handwriting by eye. Schema Law reads uploaded PDF, PNG and JPEG pages into text and tables and exports Word, searchable PDF and other formats. Its homepage says recognition is strongest on clean print and weaker on handwriting, and lists four commitments about how documents are handled, among them that documents are never used to train an AI model. It is not a tool for training models on old hands.