# Which AI Actually Reads Your German Church Record?

Every tool says it handles German. Nobody has published a test on an actual Kirchenbuch page, or on a Latin parish register, and the accuracy figures on the landing pages have no methodology behind them. Here is what each tool's own model catalogue shows, which model to run, and why FamilySearch's full-text search will not do this for you yet.


You did the hard part. You worked back through the census returns, found the naturalisation papers, and [got a place name out of them](/blog/find-immigrant-ancestor-town-of-origin). The parish is digitised. You have the page open, and there in the middle of it is your great-great-grandfather's baptism, written in a script that looks like a picket fence.

Now you go looking for something that will read it, and every tool you find says it handles German.

They cannot all be right, and it turns out nobody has checked. There is no published test of any of these tools on an actual German church book. There is no published test on a Latin parish register either. Every multi-tool comparison in this field, including the good ones, was run on English documents, and most of them say so in a line that gets dropped when the results are quoted.

So this is a different kind of comparison. Not a bake-off we ran, but an audit of what each tool's own catalogue and documentation actually say, checked against the marketing, all of it observed on 8 September 2026.

**The short answer.** Transkribus has the only public models built specifically for Kurrent and Sütterlin, but it will not tell you which of its models to run, and its newest one quietly supersedes its German specialist without any page saying so. FamilySearch Full-Text Search, which most people assume covers German, appears still to be English, Spanish and Portuguese only. Leo's German support is a claim about the language, not the script. Two other vendors publish round accuracy numbers with nothing behind them. And the Latin in a Catholic register fails differently from the German in a Lutheran one, which nobody writes about at all.

## First, the thing most people get wrong

FamilySearch Full-Text Search is the best thing to happen to genealogy in a decade. It left FamilySearch Labs on 30 August 2025, it covers on the order of two billion record images, it is free, and it reads handwriting well enough to find a name in a document nobody indexed.

It also, as far as we can establish, does not read German.

FamilySearch's own description puts the live languages at English, Spanish and Portuguese, and names Chinese, French, German, Dutch and Italian as planned additions. That framing has been in place since late 2025 and we could find no announcement since, including from RootsTech this year, confirming that German, French, Dutch or Italian have shipped. That is an absence of evidence rather than a confirmed negative, so check in the product before you rely on either answer. But if you have been assuming that FamilySearch would eventually surface your German parish entry in a search box and you could stop worrying about transcription, that assumption is not yet safe.

It is also the reason this question exists at all. For English-language research the tooling question is increasingly moot, because full-text search is closing it. For German, Latin, Italian and French records, you still need something that reads the page you are holding.

## What Transkribus actually has, and what it will not tell you

Transkribus is the specialist platform, it publishes its model catalogue openly, and that catalogue is the most informative document in this field. On the date we checked it held 430 public models, 329 of them trained on handwriting.

German is the best-served language in the entire catalogue by a wide margin, with 71 models carrying a German tag. Latin has 40. For comparison, the whole of Hebrew has six. If your ancestry is German, Austrian or Swiss, you are researching in the best-resourced corner of this field, which is a genuinely good position to be in.

The problem is the opposite of scarcity. It is that seventy-one models is not a menu, it is a maze, and Transkribus does not hand you a map.

Two models matter for most people:

| Model | Covers | Training | Published CER | Released |
|---|---|---|---|---|
| German Genius (Super Model) | German: Kurrent, Sütterlin, Fraktur and modern type, 16th to 20th century | 21.1 million words | 4.5% | 14 January 2025 |
| Text Titan II | Twelve Latin-script languages including German and Latin, print and handwriting | 250 million words | not published | 2 June 2026 |

Here is what nobody has written down. Transkribus's announcement of Text Titan II states that it reduces character error rate by roughly 47 per cent on average against its predecessor, that error rates on handwritten material are roughly halved, and that it "outperforms previous language-specific Super Models." German Genius is a previous language-specific Super Model. So Transkribus is telling you, in a blog post, that its new general model beats its German specialist.

It does not tell you this anywhere you would look. The German Genius model page does not mention it. The Kurrent landing page does not name either model. Neither does the Sütterlin page or the old German scripts page. And the Text Titan II announcement never uses the words Kurrent or Sütterlin at all, which means a researcher searching for the script they are actually holding will never be shown the model they should probably be running.

**On model selection, the documentation says to not worry about it.** The old German scripts page states that the system detects and handles different script types automatically and handles documents mixing Kurrent, Sütterlin and Fraktur without manual model selection. That may well be true and it is a reasonable default. But it means that if you want to know whether German Genius or Text Titan II reads your 1782 village register better, the vendor's position is that you should not need to ask, and there is nowhere to find out.

**On accuracy, the landing pages and the model cards disagree.** The Kurrent, Sütterlin and old German scripts pages all carry the same claim of 95 to 99 per cent character accuracy, copied verbatim between them, with no methodology, no source and no date. The church records page gives a different figure, above 95 per cent character accuracy on well-preserved German Kurrent, which is to say under 5 per cent character error. The only German figures with any provenance are the two on the model cards: 4.5 per cent for German Genius and 8.3 per cent for the older community model German Giant I. Those are consistent with the lower claim and not with the upper one. Treat 99 per cent as marketing and 4.5 per cent as the number with a document behind it.

There is one more thing worth knowing, and it is a small mark against a company we otherwise rate. Transkribus's Sütterlin page states that the Nazi regime abolished German cursive scripts in 1941 in the same breath as the Fraktur typeface ban. Those were two different acts. Martin Bormann's circular of 3 January 1941 abolished broken typefaces in print. A separate circular of 1 September 1941 ended the teaching of Kurrent in schools, and from 1942 Latin script became the school standard. It is a minor error, but it is on the page that is supposed to establish the vendor's authority on the script, and it is the kind of thing worth noticing about any source.

## The other tools, briefly

**Handwriting OCR** names Sütterlin, Kurrent and Gothic hands explicitly on its genealogy page, alongside German, French, Latin and English. That is a script-level claim, which is more than most make. It publishes no accuracy figure of any kind for them, no model name and no test. Its dedicated page on Latin manuscript transcription, which search engines still list, now returns a not-found error.

**handwritingtotext.org** has pages for Kurrent, Sütterlin, old German handwriting and Latin manuscripts, each claiming 98 per cent accuracy. No model is named anywhere, no methodology is given, and the free allowance is described only as a small one-time credit balance without saying how much. A round number with nothing behind it is not a weaker claim than 95 per cent. It is a different kind of statement.

**Leo** is the interesting case, because its German claim is narrower than it looks. Its own material says the model works very well for texts in Latin, French, German, Spanish and other major European languages. That is a claim about languages. Kurrent and Sütterlin are never named, and they are scripts. A model can be excellent at nineteenth-century German written in Latin cursive and still be defeated by the same language in Kurrent, because the difficulty is in the letterforms rather than the vocabulary. Leo is also candid about the edges, saying plainly that it struggles with less common hands and is still improving support for older scripts.

Where Leo does make a specific claim is English secretary hand, where it reports roughly 5 per cent character error and 61 per cent fewer errors than the next best model tested. The competitors are not named and the test set is not published, so treat it as a vendor claim, but it is at least a claim precise enough to be wrong, which is more than the round numbers offer.

**MyHeritage Scribe AI** launched on 4 March 2026 and says that all languages supported on the MyHeritage website are supported for uploaded documents. That is circular. A website interface language list is not a statement about handwriting models, and no script is named. Independent testers have run it on German and Italian records with reasonable results, and the most useful criticism of it so far is not about accuracy at all: there is no way to correct what it produces.

**Ancestry's document transcription** covers German among other languages, but the constraints matter more than the coverage. JPG and PNG only, no PDF. One document at a time. It works on documents attached to public trees, and it requires a World Explorer subscription or above. For a folder of photographs you took yourself, that is a set of limits worth knowing before you plan around it.

## Nobody has tested any of this on a German document

This is the finding that surprised us most, and it holds up to a fairly determined search.

Every independent, non-vendor comparison of these tools was run on English. The most careful of them, from a well-regarded genealogy educator, tested six tools on a single 1792 Virginia deed and says outright that the tests are always in English from the seventeen and eighteen hundreds. One independent test covers Italian and English and found the general language model beat the specialist on Italian while losing to it on English, which is a genuinely interesting result and a sample of one document per language. The vendor comparisons benchmark themselves and use English test sets.

There is no published comparison of any two tools on a Kurrent church book. There is none on a Latin baptism register. For the two document types that account for an enormous share of American genealogy, the evidence base is empty, and the confident rankings circulating in this space are all measurements of something else.

We are not exempt from that. We have not published such a test either.

## Catholic registers fail differently, and this is not written about anywhere

If your family was Catholic, in Germany, Ireland, Italy, France or the Habsburg lands, your parish register is probably in Latin, and Latin creates a specific problem that has nothing to do with how hard the handwriting is.

Latin parish entries are templates. A baptism gives the date, the child, the parents with the child marked as their legitimate issue, and the godparents, in nearly the same words every time for two hundred years. A marriage says that these two were joined in matrimony, and names the witnesses. A burial is thinner still, often little more than a name, *mortuus* and *sepultus*. The Irish genealogy guides put it plainly: translating the Latin is not difficult, because the entries follow the same format over and over.

Now think about what that means for a machine.

Perhaps eighty per cent of the characters on the page are that template. The template is the easy part for any recognition system, and a language model will reproduce it almost perfectly whether or not it can read the ink, because it has seen the formula thousands of times. The remaining twenty per cent is the entire reason you opened the page: the surname, the given names, the village, the day, the year, the godparents.

Those are the tokens with no linguistic context to recover them from. And the two classes of tool fail on them in opposite ways. A character-level recognition model, faced with a name it cannot resolve, tends to produce visible nonsense, and nonsense announces itself. A language model, faced with the same ink, produces a real surname of the right nationality for the region, correctly spelled, sitting comfortably in a well-formed Latin sentence. There is no visual signal that this happened.

Which means the more formulaic your record type, the more flattering the headline accuracy and the less that accuracy tells you. A Latin parish register is close to the worst case for that mismatch, and it is the single most common document in Catholic European genealogy. We wrote about the general form of this problem in [how accurate is AI at reading old handwriting](/blog/how-accurate-is-ai-handwriting-transcription); the Latin register is where it bites hardest.

Transkribus does address parish-register Latin on its church records page, mentioning the abbreviations and ligatures found in registers and pages that mix Latin headings with Kurrent body text. Its dedicated Latin page, though, is oriented to medieval manuscripts, to Caroline minuscule and Gothic textura and Beneventan. That is a different discipline from a village priest's eighteenth-century hand, and it means the page you would naturally land on is not about your document.

## If your records are English

Secretary hand, the dominant English hand from roughly 1500 to 1700, has a property worth understanding before you judge any tool on it.

Its abbreviations do not compress letters, they delete them. A macron over a vowel marks an omitted nasal consonant. Brevigraphs stand in for whole words or syllables. The information is not present in the strokes at all and has to be reconstructed from convention.

That is a structural limit rather than a quality problem. A character-level model cannot recover a letter that was never written, so it will faithfully transcribe an incomplete word. A language model can expand the abbreviation correctly, and can equally well expand it into something plausible that the scribe did not mean. Neither behaviour is a bug. They are different tools doing what they are for, and knowing which one you are holding tells you what to check.

For most American researchers with British ancestry this matters less than it sounds, because parish registers after 1700 are in ordinary hands and the pre-1700 material is largely wills, deeds and manorial records. But if you have got back that far, the tool question changes shape.

## How to test any of them on your own record in twenty minutes

Since the comparison does not exist, run a small one. Free tiers exist for this and it costs nothing.

1. **Pick three pages, not one.** One clean and typical, one where the ink has failed or the hand is cramped, and one that mixes languages, such as a Catholic register with Latin headings and German entries. A single easy page makes every tool look excellent.
2. **Write out the eight fields you actually need, yourself, first.** Surname, given names, day, month, year, place, age, occupation. Do this before you run anything, or you will unconsciously grade the machine against its own answer.
3. **Run all three pages through two tools.** Where two independent systems agree, confidence rises sharply. Where they diverge, they have found the hard words for you.
4. **Score the fields, not the page.** A tool that renders the Latin formula beautifully and invents a godparent has failed at the only job you gave it.
5. **On a Transkribus run, try both German Genius and Text Titan II on the same page** and see whether the newer general model really does beat the German specialist on your hand. Nobody has published this comparison, and on your own corpus it is a ten-minute question to answer.
6. **Check what happens to the parts that were never written.** In secretary hand, look at whether abbreviations were expanded, and whether the expansion is marked as an expansion or presented as if it were on the page.

Twenty minutes of this tells you more about your records than every published benchmark combined, because every published benchmark was run on somebody else's.

## What KleioBase does, and what it does not

We are one of the tools in this category, so read this as disclosure.

KleioBase uses Google Gemini for extraction. It reads [old German scripts](/transcribe/german-records) including Kurrent, Sütterlin and Fraktur, and [Latin parish registers](/transcribe/latin-parish-records), and it handles pages that mix the two, which German Catholic registers routinely do. The transcription is produced in the document's original language and script and is never translated, with a translation and structured fields alongside it, so you are never left holding only the English. Before processing you can add context, meaning the language, the place and the approximate year, and draw region boxes to point at the one entry you want on a page holding eight. For difficult hands there is a deep scan option that runs a stronger model at a higher credit cost. Nothing enters your knowledge base automatically: every record passes through a Review state where each field is editable beside the image, and confirming is what creates profiles. The mechanics are in [uploading records](/docs/uploading-records).

The honest part.

**We have not published a test on a Kirchenbuch either, and we do not publish an accuracy figure.** This article criticises an industry for asserting German coverage without evidence, and we assert German coverage. Our reasons for declining to publish a single headline number are set out in our [article on accuracy](/blog/how-accurate-is-ai-handwriting-transcription), and they are honest reasons, but they leave us in the same position as everyone else on this specific question.

**A general model is a different bet from a trained one.** If you have one clerk's register and several hundred pages of it, training a private Transkribus model on your own corrected pages will beat any general model on that corpus, and that is the right route for a serious single-parish project. We are built for the opposite case, where you have forty documents from six parishes in three languages and training a model on any of them is not worth the effort.

**Silent modernisation is a risk we carry too.** A general-purpose model that regularises an archaic spelling or expands a Latin abbreviation without telling you is a real hazard, and the control for it is the review step and your own eyes on the image, not our assurance.

## The question to ask any vendor

Not "does it read German." Every one of them says yes, and none of them are lying, because German covers a modern typed letter and an 1806 Kurrent burial entry equally.

Ask instead: **on what script, from what century, was that number measured?** Nobody in this industry can currently answer that with a German church book or a Latin parish register, which tells you something about the state of the evidence and nothing about whether the tool will work on your page.

Your page is the only benchmark that matters. It takes twenty minutes and three documents to run it.

For learning to read the script yourself, which is worth doing for a few treasured documents whatever tool you use, see our guide to [Kurrent and Sütterlin](/blog/reading-old-german-records-kurrent-sutterlin).

Canonical: https://kleiobase.com/blog/which-ai-reads-german-church-records-kurrent-latin