Handwritten Text Recognition (HTR): A Practical Guide

This is an introduction to handwritten text recognition (HTR) for anyone meeting the term for the first time — what it is, where it sits beside OCR, and how to fit machine transcription into serious work on historical documents. It matters because a tool that reads a page for you can also quietly rewrite your sources, and knowing where the line falls is the difference between a first draft you can trust and one you cannot.

Handwritten text recognition is the automated transcription of handwriting from page images into machine-readable text. Unlike classical OCR, which classifies printed glyphs one at a time, HTR reads a whole line of connected script in context, because cursive has no reliable letter boundaries to segment. Modern HTR uses sequence models — recurrent networks, then attention, now transformers — trained on real handwriting rather than typefaces; a separate guide explains what an HTR model is and how it reads a line, stage by stage. On difficult historical hands it works well enough to be useful. But it is not solved: even strong models leave errors a reader must check against the page. For the full mechanical contrast — where each approach breaks on historical print and cursive, and how ATR blurs the labels — see HTR vs OCR.

That last sentence is the one most guides skip, and it is the one that should govern how you use the technology. This piece explains what HTR is, the main kinds of model you can use today, how to read the accuracy numbers you will encounter, and how to fit machine transcription into serious work without letting it rewrite your sources. It sits under our broader guide to handwritten text recognition as an entry point for readers meeting the term for the first time.

Script, language, and hand: what "supported" actually means

Ask "can HTR read my documents?" and the honest answer has three parts, not one.

Script

Script is the writing system — the alphabet family. Most historical HTR tooling is built for the Latin script: the alphabet used to write English, French, German, Dutch, Spanish, Italian, Portuguese, Polish, and dozens of other languages, including Latin itself. Non-Latin scripts — Greek, Cyrillic, Hebrew, Arabic, and the Indic and East Asian systems — are a genuinely different problem, and support for them lags the Latin-script tooling by years. A vendor claim of "100+ languages" is really a claim about script coverage plus language priors. It does not mean every hand in every language reads equally well. (link unchanged: ilc.csic.es — see Links)

Language

Language shapes the priors a model brings to ambiguous marks, and coverage is uneven. Published benchmarks over-represent English. Peer-reviewed evaluation is thinner on German Kurrent and Sütterlin, French notarial hands, and Spanish and Italian cursive. A model strong on English can be mediocre on a German parish register.

Hand

Hand is the individual style of script — secretary hand, court hand, italic, Kurrent. This is often the real determinant of success, and the one most easily hidden by a headline figure; we look separately at whether AI can read cursive, where modern and historical hands part company. Our guide to which languages and scripts HTR can read works through why the honest unit of support is script + language + hand together, not a language count.

How to read an accuracy number

HTR quality is reported as character error rate (CER): the proportion of characters wrong through insertion, deletion, or substitution, borrowed from speech recognition. A perfect transcription is 0% CER. On difficult historical material, anything under about 10% is respectable, and figures above that are not unusual. Word error rate (WER) does the same at the word level and always runs higher, because a single wrong character spoils a whole word.

The trap is treating one number as a property of a tool. It is not. Published CERs are almost impossible to compare across studies, because they come from different datasets, different evaluation protocols (strict versus lenient), different page splits, and different amounts of language-model post-processing. Romein and colleagues showed the effect directly in 2025: testing five HTR engines across datasets in several scripts, they found the engine that did best out of the box on Latin-script material was not the one that did best on other scripts once trained. A model that wins on one collection can lose on another. And the organizers of the ICDAR 2024 competition on handwritten historical ciphers were blunt that performance on the hardest material remains far from satisfactory.

Two practical consequences follow. First, a CER measured on someone else's collection tells you little about yours; the only number that matters is the one you measure on a sample of your own hands. Second, decide your tolerance before you start. Roughly 5% CER is publishable-quality first-pass text; 15% is usually still good enough to make a collection searchable but not to quote from without close correction.

Approaches to HTR today: specialist models, general AI models, and OCR

There are now four broad ways to get a machine reading of a historical page, and they fail in different ways. This section compares approaches; if you are choosing between products, our comparison of HTR software takes them one by one.

Training a model for one collection

For years, using HTR meant training a model. You transcribed a few thousand words of ground truth in a page editor, aligned it line by line, and built a model specific to your hand and genre. Transkribus still recommends around 10,000 words per hand for cursive to get good results. That is a serious investment before you transcribe a single page you actually care about, and it remains the right path for a large, homogeneous collection in an unusual hand. (We weigh up training your own model in a separate guide.)

Ready-made specialist models

The alternative is a ready-made specialist model: one pre-trained on a broad range of historical material, which you can point at a page with no per-collection training. Leo's own model, ATR-1, is built for exactly this. It is a transformer model tuned specifically for historical documents in the Latin alphabet, handwritten and printed: secretary hand, court hand and italic, and the early-modern founts that conventional OCR mishandles. It works out of the box, with no ground-truth step and no training on your hand.

Three things set it apart from a general-purpose model.

It keeps the page's own text. ATR-1 is trained to transcribe what is on the page rather than normalize it. It does not over-correct or silently modernize, so an archaic spelling, a strikethrough, a marginal addition, a macron or a long s survives into the output instead of being quietly tidied away. For a historian, that is the difference between a transcription and a paraphrase.
It reads the image at high resolution. General models downsample the images you give them. ATR-1 weighs high-resolution visual evidence against context, the way a trained reader uses the rest of a word to settle an unclear letter.
It watches for its own failures. ATR-1 checks its output for failure patterns, such as a line repeated over and over. Output that looks wrong is hidden, retried automatically, and refunded if it still cannot succeed. The errors that do get through tend to be the recoverable kind: a wrong letter or word you can check against the image shown beside the text, not fluent invented prose.

Corrections made in Leo feed back into training, so the model improves with each release. Other ready-made models exist too: Transkribus offers its Text Titan models for Latin scripts, and the open-source ARletta models cover historic Dutch. "Zero-shot" means usable without fine-tuning, not that fine-tuning never helps; for the hardest hands it still can. But for most researchers, most of the time, a good specialist model removes the training bottleneck.

Conventional OCR, and why historical print trips it

Conventional OCR engines remain the right tool for clean modern print, and on a clean nineteenth-century page they can do well. But they were built and trained for modern, standardized type, and accuracy falls sharply on older print even when a person reads it easily (Springmann & Lüdeling, 2017; Cordell, 2019). The causes are specific. The long s (ſ) is read as an f: in a study of an eighteenth-century collection, words containing "f" made up 11% of a hand-keyed corpus but 33.7% of the OCR'd one (Hill & Hengchen, 2019). Ligatures, blackletter and Fraktur score lower; historical spelling and abbreviation get "corrected" by the engine's built-in dictionary; and on multi-column newspapers OCR can read most of the words and still put them in the wrong order. In a 2026 benchmark of English print from 1612–1807, the best vision-language models made about one character error in forty, against about one in five for Tesseract (Clifford et al., 2026). ATR-1 reads this kind of print as well as handwriting, and its source-integrity rule applies to both: the long s stays a long s, and the archaic spelling survives.

General vision-language models (Claude, GPT, Gemini)

The large general models from Anthropic, OpenAI and Google now read images, and on handwriting they have improved quickly; on English handwriting of the 18th and 19th centuries, published tests put Google's Gemini ahead of the other frontier models (November 2025). They are useful, but they are generalists. They downsample the image, lean heavily on what is linguistically probable, and so tend to quietly normalize: a more familiar spelling in place of the one on the page, or tidied punctuation and capitalization. On a source you intend to quote, that is the error that matters most, and it is the one ATR-1 is trained to avoid. For more, see can AI read old handwriting and our assessment of whether Gemini can read handwriting.

Choosing a model per document in Leo

You don't have to pick one approach for a whole archive. Leo has a model library: in the transcription window, one menu lets you choose the model for each document. ATR-1 is the default, with more than thirty general models from OpenAI, Google Gemini and Anthropic Claude and cloud OCR engines from Google, Microsoft and Amazon alongside it, for comparison or for cheap clean print. Prices are per image and vary by model: ATR-1 is always 1 credit an image, while a small model such as GPT-6 Luna is 0.02 credits. A rough guide:

Handwriting: ATR-1, which is specialized for historical hands and keeps the page's spelling; a general model as a second reading if you want one
Early-modern print, Fraktur, newspapers, or anything you will quote diplomatically: ATR-1
A clean, legible printed page, such as a 19th-century book or a typescript, where you just want the text: a small model such as GPT-6 Luna, for a fraction of a credit
A hard reading you want to check: run a second model; Leo keeps each run in its own tab, so you can compare them line by line

In our tests, the smallest models were as good as the alternatives on clean books and typescripts, but tended to modernize early-modern spelling and could stop partway through a dense newspaper page without saying so.

The "just paste it into ChatGPT" problem

The most common thing a newcomer reaches for is a general chatbot: photograph the page, upload it, ask for a transcription. It is worth understanding precisely why this is risky, because the failure mode is subtle.

General multimodal LLMs weight linguistic plausibility heavily and downsample the images you give them. On a difficult hand, that combination produces fluent errors rather than recoverable ones. A specialist HTR model that misreads tends to produce garble you can spot — a nonsense character, a broken word. A general LLM tends to produce a clean, confident, grammatical reading that happens to be wrong, and may insert plausible archaic spellings the source never contained. Evaluations of these pipelines document exactly this "over-historicization" and silent normalization. The models have improved fast, but improvement has not changed the kind of error that survives: the fluent, plausible misreading. That is why ATR-1 is built the other way round, reading the image at high resolution and transcribing what is there, so that its mistakes are the kind you can see and check.

The deeper issue is that a language model improves average word plausibility and can therefore overwrite a genuinely preserved archaic spelling or abbreviation — the double-edged role of language priors, with no agreed fix. Our longer treatments of why ChatGPT struggles on historical handwriting and of fluent-but-wrong LLM transcription errors work through the mechanism in detail. The short version: a wrong transcription that looks right is more dangerous than one that looks broken, because you will not think to check it.

Where crowdsourcing still wins — and where it doesn't

For decades the default for large backlogs has been human volunteers. It is highly accurate. The Smithsonian Transcription Center has run since 2013; climate-data-rescue projects reach about 99.5% accuracy on tens of thousands of weather values through replicate consensus; the Transcribe Bentham analysis found only around 1% of words needed staff alteration. Crowdsourcing produces excellent text.

What it cannot easily do is scale. It is bounded by volunteer supply, training, and reviewer fatigue — that same Bentham study clocked an average of 207 seconds of checking per transcript. Past a few tens of thousands of pages, recruitment becomes the constraint, which is why large projects increasingly run a hybrid: machine transcription for the first pass, human effort concentrated on verification and the hardest pages. That division of labor — machine reads, human checks — is the shape most serious HTR workflows now take.

Fitting HTR into real work

A defensible workflow treats machine transcription as a first draft, never a finished text (our guide to HTR for historians goes further into research practice). In practice that means a few disciplines.

Keep interpretation separate from the base transcription. The faithful reading of the page is one artifact; a modernized, translated, or corrected version is another, and conflating them corrupts your source. Verify against the image, not against the plausibility of the output — and prioritize the high-stakes tokens: names, dates, numbers, sums, and boundaries, where a single misread character does real damage. And keep your paleography sharp, because the skill that once read the page now verifies the machine; a model that hands you fluent output is precisely the one that needs a reader who can tell when it is fluently wrong. If you want to learn to read the hands yourself, Leo Academy's free secretary hand course covers the everyday hand of Tudor and Stuart parish registers, wills and inventories.

Where HTR ends, other stages begin. Transcription sits after capture — after you have photographed the documents well enough to read, and, where the ink has faded, worked out whether you have a capture problem or a recognition problem (see can AI read faded documents) — and before the metadata and publishing standards that make a collection findable and preservable. No single tool spans all of it, and a guide that tells you otherwise is selling something.

Handwritten text recognition has reached the point where it changes what a lone researcher or a small archive can attempt: collections that would have stayed unread in a lifetime of hand-transcription become searchable in weeks. But the technology reads; it does not understand, and it does not know when it is wrong. The judgment about what a document actually says — and what it means — still rests with the reader who knows the hand, the period, and the language. The machine gives you a faster first pass. The scholarship is still yours.

The guides in this section

HTR vs OCR: What Matters for Historical Documents — the full comparison: where each approach breaks on historical print and cursive, and what to ask before choosing a tool
How Does HTR Work? — the pipeline stage by stage, from image capture to export, and the model architectures that do the reading
HTR for Historians — the classes of tool a historian chooses between, and how to keep source integrity through the workflow
Can AI Read Old Handwriting? — what current models manage, where they fail, and which errors survive proofreading
Can AI Read Cursive? — why modern cursive is easy and secretary, Kurrent and procesal hands are not
Can AI Read Faded Documents? — a triage guide to faded ink, foxing and show-through: capture, recognition or conservation problem
HTR Supported Languages — why script comes first, then language, then the hand on the page
How to Test AI Transcription on Your Own Documents — an afternoon protocol: sample, key a reference, score CER and WER, read the errors
OCR for Old Books — why historical print breaks standard OCR, and which model to use for each kind of book
Fraktur OCR — reading and converting German blackletter print, and the letters that trip OCR

Frequently Asked Questions

What is handwritten text recognition?

Handwritten text recognition (HTR) is the automated transcription of handwriting from page images into machine-readable text. Classical OCR was built for machine print and classifies glyphs one at a time; HTR treats a whole line as a sequence because cursive has no clean letter boundaries. That is enough to orient the term — the full contrast, including historical print, ATR, and how to read accuracy claims, is in HTR vs OCR.

What are the main approaches to handwritten text recognition today?

There are four broad ways to get a machine reading of a historical page, and they fail in different ways. You can train a model on your own ground truth for one collection, which repays the effort for a large, homogeneous collection in an unusual hand. You can use a ready-made specialist model pre-trained on historical documents, such as Leo's ATR-1, which reads with no per-collection training and keeps the page's own spelling. Conventional OCR remains the right tool for clean modern print, but its accuracy falls sharply on older type. And general vision-language models such as Claude, GPT and Gemini read many hands well but tend to quietly normalize spelling. In Leo you choose the model for each document from one menu.

What is a good character error rate for handwritten text recognition?

On difficult historical material, anything under about 10% character error rate (CER) is respectable, and figures above that are not unusual; a perfect transcription is 0%. Roughly 5% CER is publishable-quality first-pass text, while 15% is usually enough to make a collection searchable but not to quote from without close correction. The trap is treating one number as a fixed property of a tool. Published CERs are almost impossible to compare across studies, because they come from different datasets, evaluation protocols, page splits, and amounts of language-model post-processing. The only number that truly matters is the one you measure on a sample of your own hands.

Can I just paste a handwritten document into ChatGPT to transcribe it?

You can, but check the result with care. General AI models have improved quickly, but their failure mode is subtle. They weight linguistic plausibility heavily and downsample the images you upload, so on a difficult hand they produce fluent errors rather than recoverable ones: a clean, confident, grammatical reading that happens to be wrong, or a modern spelling quietly put in place of the one on the page. A wrong transcription that looks right is more dangerous than one that looks broken, because you will not think to check it. A specialist model such as Leo's ATR-1 is trained to transcribe what is on the page rather than normalize it, reads the image at high resolution, and shows its transcription beside the image so you can check it.

What does it mean when an HTR tool claims to support "100+ languages"?

A claim of "100+ languages" is really a claim about script coverage plus language priors, not a promise that every hand in every language reads equally well. Honest support has three parts. Script is the writing system: most historical tooling is built for the Latin alphabet, while Greek, Cyrillic, Hebrew, Arabic, and Indic and East Asian systems lag it by years. Language shapes the priors a model brings to ambiguous marks, and coverage is uneven, with English over-represented in benchmarks. Hand is the individual style — secretary, court, italic, Kurrent — and is often the real determinant of success. The honest unit is script, language, and hand together.

Which approach should I use for my documents?

It depends on the document. For handwriting, start with a specialist model trained on historical hands, such as Leo's ATR-1, which keeps the page's own spelling instead of quietly modernizing it. ATR-1 is also the choice for early-modern print, Fraktur or newspapers. For a clean, legible printed book or typescript, a small, cheap model or an OCR engine is usually enough. Leo lets you choose the model per document from one menu and keeps each run in its own tab, so you can compare two readings of a hard page. Whatever you choose, test it on a sample of your own documents before you commit a collection to it.