This is a guide to reading manuscript abbreviations — the suspensions, contractions, brevigraphs, superscripts, and ligatures that scribes used for a thousand years. It sets out the five classes to learn, the high-frequency marks that carry most of the weight, and the editorial choice every transcriber faces: whether to preserve an abbreviation or expand it. If old handwriting is slowing your research, knowing this system is where the work becomes tractable.

Manuscript abbreviations are not random shorthand. They form a rule-governed system, inherited from Roman practice and refined across a thousand years of scribal work, and once you know the categories they fall into, most of them resolve on sight. There are five classes to learn — suspension, contraction, brevigraph, superscript, and ligature — plus a handful of high-frequency marks that account for the bulk of what you will meet on a page. Learn those first, and the rest becomes a reference-lookup problem rather than a wall.

This guide sets out the typology, works through the marks you will encounter most often across English, French, Spanish, German, Dutch, and Italian records, and closes with the question every scholar eventually faces: whether to preserve an abbreviation or expand it, and what that decision costs. It sits alongside our broader work on reading old handwriting and paleography, and assumes you can already make out the letterforms — the problem here is what to do when the letters aren't all there.

Parchment was expensive, ink was slow, and hands cramped. Abbreviation saved all three. But it was never improvised. The system descends in a direct line from Tironian notes — the shorthand devised around 63 BC by Marcus Tullius Tiro, Cicero's freedman and secretary, which began with roughly 4,000 signs and grew to some 14,000 by the Carolingian period. One Tironian sign, the et symbol (⁊), still sits in Unicode today. Christian scribes layered their own conventions on top: the nomina sacra, contractions of holy names marked with an overline, of which around fifteen forms are attested in Greek New Testament manuscripts by the fourth century.

The practical consequence for a reader is this: abbreviation is conventional, so it is learnable. A scribe writing a will in 1590 and a notary drawing a contract in 1720 drew on the same shared repertoire, with local accents. And it did not die with the Middle Ages. Abbreviation persisted in notarial, ecclesiastical, and bureaucratic writing well into the eighteenth and nineteenth centuries, and German Kurrent chancery documents carried it into the early twentieth.

Every mark you meet belongs to one of five structural types. This typology is standard across the modern handbooks — Cappelli, Clemens and Graham, and Honkapohja's survey for VARIENG all use it.

A word truncated at the end, with the missing letters supplied by context. Dr for doctor. Spanish q̃ for que. The reader restores the ending. Suspension is the oldest and simplest device, and in administrative records it is everywhere: a clerk who wrote the same title fifty times a day had no reason to spell it out.

Letters omitted from the interior of a word, usually with the first and last letters retained and a mark signalling the gap. Latin dns for dominus. English y with a small superscript e for the, where the y is not a y at all but a survival of the thorn (þ). Contraction is the class that most often misleads a modern eye, because the surviving letters can look like a complete, if strange, word.

A single sign standing for a letter, a syllable, or a whole word. These are the marks that most reward memorisation, because they are frequent, distinctive, and not derivable from the letters around them. The core set:

The p-with-stroke family. A p crossed or looped in different ways yields per, par, pro, and prae — the position and shape of the stroke tells you which. Pervasive in legal and notarial records in every language.

The 9-shaped sign for terminal -us or -um, written like a small figure nine at the end of a word.

The ꝫ sign for -rum, common in genitive plurals.

The -que sign and the Tironian et (⁊) for and.

A small raised letter, often a vowel and often the first of an omitted syllable, with the rest inferred. The English wch (which) and wth (with) work this way, as does the superscript r in yor for your. In German the same principle produces raised-letter contractions throughout Kanzlei writing.

Two or more letters joined in one pen-stroke: æ, œ, ct, st, fi, ff, fl. Here a caution matters. Most ligatures are calligraphic — a matter of speed and pen economy, not abbreviation — and carry no omitted material. Only some function as genuine abbreviations. Treat the two cases differently: a ct ligature is simply ct written elegantly, while a macron over a vowel is a true abbreviation signalling an omitted nasal.

If you learn nothing else, learn these, because they recur across languages and centuries.

The macron (or titulus). A horizontal stroke over a vowel signals an omitted m or n following that vowel. Comūne for commune; German genom̄en for genommen, where the Nasalstrich stands for the missing nasal. This one mark resolves a large share of contractions in every Latin-alphabet tradition.

The long s (ſ). Not a diacritic and not a mistake, but a normal letterform used medially and initially from antiquity to about 1800 in English, and later still in German Fraktur. It is a genuine s. Its resemblance to f is the single most common misreading a modern eye — and, as we will see, a machine — makes on early-modern material.

The p-with-stroke family and the 9-sign. Between them these cover a great deal of the brevigraph traffic in legal and administrative records.

How the same system looks across languages

The repertoire is Latin-derived and shared, but each tradition inflects it. This is where a general handbook stops being enough and a language-specific tutorial earns its place.

English secretary hand (16th–18th c.) leans on thorn-based contractions — ye (the), yt (that), yor (your) — alongside wch, wth, terminal flourishes for -es, -is, and -us, and the p-with-stroke family in legal instruments. The Folger Shakespeare Library's alphabet book is a good visual primer.

French notarial and chancery hands standardised abbreviations for que, par, per, pour, avec, dict, and titles, together with the 9-shaped brevigraphs. The Newberry Library's French paleography resource catalogues the frequent forms; Prou's Manuel de paléographie latine et française (1924) remains the reference grammar.

Spanish escribano records (16th–18th c.) are usually taught in six categories — contracciones, sobreposiciones, suspensiones, siglas, convencionalismos, and entretejido. Watch for q̃ (que), p̃ or p̄ (per/por), ñ (en), and x̃ (Christo). Kasten and Needham's Syllables, Signs, Letters (2011) is the modern handbook.

German Kurrent and Kanzlei (14th–20th c.) run a heavy contraction system: er- and re- contractions, the Nasalstrich over vowels, and terminal abbreviations, documented at Ad fontes.

Dutch inherits the same Latin repertoire with regional notarial variations, though English-language scholarship on it is thin and scattered across archive guides. Italian likewise takes the Latin system with local idiosyncrasies; Cappelli's Lexicon Abbreviaturarum remains indispensable, and is freely searchable online through Ad fontes.

One honest complication for vernacular work: the line between a deliberate abbreviation and a variant spelling is genuinely fuzzy, and no consensus rule tells you which is which. A shortened form in a Dutch register may be a true suspension or simply how the writer spelled the word. Judgement, not lookup, settles these.

Here is the decision that outlasts every reading question. When you transcribe an abbreviated word, do you write what is on the page — the q̃, the macron, the 9-sign — or do you silently expand it to que, commune, -us?

The answer depends on what the transcription is for, and it must be declared and consistent. Silent expansion is inappropriate for critical editions and for any source-preserving work, because it destroys the very evidence the transcription exists to record. Abbreviation practice is itself data: it dates documents, localises them, identifies scribal hands, and distinguishes genres. Legal records and charters conventionally preserve every suspension mark for diplomatic reasons; vernacular parish registers are sometimes silently expanded but should retain names as written.

The scholarly middle path is TEI's `<choice>` mechanism, which lets you encode both the abbreviated form and its expansion — `<abbr>` for what the scribe wrote, `<expan>` for your reading — with your editorial conventions stated in the front matter. Nothing is lost; both the diplomatic and the readable text survive in the same file. For the character set itself, the MUFI recommendation maps the brevigraphs, long s, and combining marks to Unicode code points, so a ꝫ stays a ꝫ rather than degrading to plain text.

Where machine transcription helps — and where it silently doesn't

Once you are reading at scale — a whole notarial series, a run of parish registers, hundreds of pages — hand-transcription becomes the bottleneck, and the sensible question is what a machine does with these marks. The answer separates cleanly into three failure modes, and the abbreviation is exactly where each tool shows its character.

Tesseract, ABBYY FineReader, Google Cloud Vision, and Amazon Textract are trained overwhelmingly on post-1800 type. They fail on abbreviations in a predictable way: the long s becomes f or s, brevigraphs are dropped or garbled, and the macron is treated as an accent on the vowel rather than as a signal of an omitted nasal. Holley's much-cited 2009 study found the best engines reached only about 78–83% character accuracy on long-s-rich nineteenth-century newspapers, with the ſ/f/s confusion dominating the error budget. Conditions have shifted since, but the structural weakness — a glyph classifier expecting modern type — has not.

Transkribus, eScriptorium, and OCR4all behave entirely according to their ground truth. If the training transcriptions preserve abbreviation marks, the model preserves them; if they expand, it expands. Transkribus instructs users to transcribe abbreviations as they appear, tag them, and add the expansion as a property. This is principled, but it puts the burden on you: you inherit whatever labelling convention the model was trained on, and typically you train or fine-tune a model to get there.

GPT, Claude, and Gemini used directly are the seductive case. On some material they read handwriting surprisingly well: Humphries et al. reported that standalone frontier models reached CER in the 5.7–7.0% range on eighteenth- and nineteenth-century English handwriting, and better still when used to post-correct HTR output — though that is one corpus, one language, and a result that still requires verification before you lean on it. Their distinctive danger is precisely with abbreviations: they tend to silently expand and normalise, "correcting" what they see without flagging it. An LLM will confidently render q̃ as que and give you no way to know it made a choice. That is the worst kind of error, because it is plausible and invisible — the opposite of a dropped brevigraph, which announces itself.

This is the crux. For abbreviation-heavy work, the failure you can catch is far safer than the failure you can't. A tool that mangles a ꝫ leaves a visible scar; a tool that quietly resolves it to -rum erases the evidence and hands you fluent, wrong text.

This is the ground Leo's ATR-1 model is built on. It is a specialist transcription model for Latin-script material — handwritten and printed, in whatever language the alphabet records, from English wills to French notarial minutes to German parish books — and it works out of the box, with no per-corpus model to train. The design commitment that matters here is source integrity: ATR-1 transcribes what is on the page rather than smoothing it into modern prose. The long s stays a long s; the macron is not silently resolved to a nasal; archaic orthography, strikethroughs, and marginal additions survive into the transcription. Where you want an expansion or a modernisation, that is a separate, explicit step in a Transform layer that writes to a new tab and leaves the faithful base reading untouched — so the editorial act is recorded, not baked in. On a randomised 97-image sample of early-modern English manuscripts from the Folger Shakespeare Library, ATR-1 scored roughly 5% character error rate at release — 61% fewer errors than the next-best model tested — with the full benchmark and per-model figures published here. No system reads abbreviations perfectly, and none should be trusted without verification; what a preservationist model buys you is that the verification is possible at all.

If your material is early-modern print — the long s, ligatures, and typographic abbreviation that conventional OCR mishandles — the same specialisation applies, and the free tier or a paid plan is the way in. For the sequence of hands, letterforms, and dating conventions that underlie all of this, our guide to early modern paleography goes deeper on the reading itself.

Start with the typology, not the individual marks: knowing whether you are looking at a suspension, a contraction, a brevigraph, a superscript, or a ligature tells you where the missing letters are, which is most of the battle. Memorise the high-frequency signs — the macron, the p-with-stroke family, the 9-sign, the long s — because they recur everywhere and repay the effort within a page or two. Keep Cappelli and the relevant language tutorial within reach for the rest. And decide, before you begin, whether you are preserving or expanding, and hold to it — because the marks a scribe made to save time are, four centuries later, some of the best evidence you have about who wrote what, where, and when.

Manuscript abbreviations are the rule-governed shorthand scribes used for a thousand years to save parchment, ink, and effort. They fall into five structural classes: suspension (a word cut short at the end, like Dr for doctor), contraction (letters dropped from the interior, like dns for dominus), brevigraph (a single sign for a letter, syllable, or word, like the 9-shaped -us mark), superscript (a small raised letter, as in wch for which), and ligature (joined letters such as æ or ct). The system descends from Roman Tironian notes and, because it is conventional, it is learnable.

What does the long s (ſ) mean in old handwriting?

The long s (ſ) is a normal letterform, not a diacritic or a mistake — it is a genuine s used medially and initially from antiquity to about 1800 in English, and later still in German Fraktur. Its resemblance to f is the single most common misreading a modern eye makes on early-modern material, and machines share the weakness: general-purpose OCR routinely renders it as f or s. It is a letterform question, not an abbreviation, but it sits alongside abbreviations because both trip up readers and tools trained on modern type.

What does a macron or line over a letter mean in manuscripts?

A macron — a horizontal stroke over a vowel, also called a titulus — signals an omitted m or n following that vowel. So comūne expands to commune, and German genom̄en to genommen, where the Nasalstrich stands for the missing nasal. This single mark resolves a large share of contractions across every Latin-alphabet tradition, which is why it repays memorising early. Note the caution: a macron over a vowel is a true abbreviation, whereas a ct or st ligature is simply elegant penmanship carrying no omitted material. Treat the two cases differently.

Should you expand or preserve abbreviations when transcribing?

It depends on what the transcription is for, and the decision must be declared and held consistently. Silent expansion is inappropriate for critical editions and source-preserving work, because abbreviation practice is itself evidence: it dates documents, localises them, identifies scribal hands, and distinguishes genres. Legal records and charters conventionally preserve every suspension mark; vernacular parish registers are sometimes expanded but should retain names as written. The scholarly middle path is TEI's `<choice>` mechanism, which encodes both the abbreviated form and its expansion in one file — nothing is lost, and both the diplomatic and readable text survive together.

Can AI or OCR transcribe manuscript abbreviations accurately?

No tool reads abbreviations perfectly, and the failure modes differ. General-purpose OCR (Tesseract, ABBYY, Google, Amazon) is trained on modern type and drops or garbles brevigraphs, mistakes the long s for f, and treats the macron as an accent. General LLMs read some handwriting well but tend to silently expand abbreviations — rendering q̃ as que with no flag — which is the worst error because it is plausible and invisible. Leo's ATR-1 model is built to preserve the marks rather than smooth them, so verification stays possible; on a Folger sample it scored roughly 5% character error rate at release.

Five classes of scribal shorthand - suspension, contraction, brevigraph, superscript, and ligature - and the high-frequency marks that resolve most abbreviations on sight.

Manuscript Abbreviations and Ligatures: A Reader's Guide

Secretary hand, court hand, archaic spelling, and contractions make older documents hard to read. Here you'll find practical guidance on deciphering historical hands.

Manuscript Abbreviations and Ligatures: A Reader's Guide to Scribal Shorthand

Secretary hand is the everyday handwriting of English-speaking Europe from roughly 1500 to 1700 — the script behind most parish registers, wills, deeds, and letters of the period. This guide explains how to read secretary hand from first principles: the letterforms, the abbreviations, and the habit of reading by context that together turn an illegible page into a legible one. If old handwriting has stalled a line of your family research, this is where you start.

Learning to read secretary hand comes down to three things: the letterforms that differ systematically from modern ones (the secretary e, c, r, h, and the long s), the abbreviations that delete whole syllables, and the habit of reading by context rather than by exact spelling. It is a learnable skill, not a gift. The difficulty is structural, not a matter of messy penmanship.

That last point is worth settling before anything else. The reason a will from 1580 stops you cold is not that the scribe wrote badly. It is that the writing system itself differs from the one you know. Once you see the differences as a system, they stop being noise and become rules — and rules can be learned. This guide walks through what those rules are, in the order a beginner actually needs them, and sits within a broader set of resources on reading old handwriting and paleography that covers other hands and periods.

What secretary hand is — and what it is not

Secretary hand developed in the early sixteenth century, descending from earlier cursive business hands, and remained the dominant everyday script through the sixteenth and seventeenth centuries before italic slowly displaced it. It was used across much of northern Europe — for English, German, Dutch, Welsh, and Gaelic writing — which matters because the same letterforms recur across languages. The skill you build on an English will transfers, in part, to a Dutch register.

Two distinctions save beginners a great deal of confusion.

First, secretary hand is not the same as court hand. Court hand is a narrower, more angular script used specifically in English legal records from the thirteenth to eighteenth centuries, descending from a different lineage — a gothic chancery hand. The two overlap in time, and courts used secretary hand too, so you will meet both. But if a document looks far more cramped and unfamiliar than the wills and letters around it, you may be looking at court hand, which is a separate study.

Second, secretary hand is overwhelmingly a vernacular script. This is the misconception that trips up newcomers most often. The documents you will read are in English (or German, or Dutch) — not Latin. Latin appears only in formulaic patches: legal boilerplate, dating clauses, standard abbreviations. Do not expect to need Latin to read a Tudor will. You need to read early modern English written in an unfamiliar hand.

By the later seventeenth century, educated writers were blending secretary and italic letterforms into what paleographers call mixed hand — the transitional stage that eventually became modern handwriting. There is no clean cut-off year. If your document sits around 1650–1700, expect a mixture.

The single most efficient thing a beginner can do is learn the handful of letters that look nothing like their modern equivalents. Most of the alphabet is close enough to read on sight. It is a small set of troublemakers that does the damage.

The usual suspects are the secretary e, c, r, and h — each of which can be startlingly unlike its modern form — along with a two-stroke or reversed e that appears where you least expect it. Rather than memorising these in the abstract, work from an alphabet chart alongside a real manuscript. The free tutorials from The National Archives, the Society of Genealogists, and Cambridge's English Handwriting 1500–1700 online course all pair a letter-by-letter chart with graded practice manuscripts and model transcriptions. Cambridge's course in particular builds through lessons of increasing difficulty, each with a manuscript to transcribe and a semi-diplomatic transcription to check yourself against — which is exactly how the skill is built, one page at a time.

The long s (ſ) is the letterform beginners most often get wrong, because it looks like an f and the mistake is easy to make.

It is a distinct graph for /s/, used at the start of a word or in the middle, but never at the end. So "yes" could be written yeſ, and "shelf" is ſhelf — but a word-final s is always the modern round s. It is also not used before f, k, or b. The visual tell, as the Society of Genealogists guide notes, is the cross-stroke: an f normally has a bar going right through the stem, while the long s has, at most, a nub on the left side that does not fully cross. It is not an old-fashioned f; it is an s.

The long s is stubborn. It survived in manuscript hands into the second half of the nineteenth century, long after it vanished from London print around 1800. You will meet it well outside the secretary-hand period.

Learn the abbreviations, because they delete letters

The second structural difficulty is that secretary-hand writers abbreviated constantly, and their abbreviations do not merely shorten words — they delete whole letters and syllables that you are expected to supply. Miss the mark and you miss the missing letters entirely.

The macron (or tilde/titulus) — a horizontal stroke over a vowel — marks an omitted nasal, m or n. So coũent stands for convent. When you see a bar floating over a vowel, suspect a dropped m or n.

Superscript letters signal contraction: raising one or more letters above the line, often with a suspension mark. w̃ch is "which," w̃th is "with," yor is "your." Cambridge's conventions page sets out how editors expand these.

Brevigraphs are single symbols standing for a whole word or syllable — the ꝑ that resolves to per, pro, or pre depending on context, or a stroke standing for terminal -us.

The Tironian et (⁊), a small mark meaning "and," survives from seventh-century shorthand and is the ancestor of the modern ampersand.

These are not decoration. They carry meaning, and a transcription that silently drops them has lost information from the page.

One abbreviation deserves its own note because it powers a famous misconception. The thorn (þ) was the Old English letter for the th sound. In many secretary-hand forms it came to look like a y, and early printers — lacking a thorn sort in their type — substituted a y. That is the entire origin of "Ye Olde." As the history of the letter thorn records, the word was always "the," pronounced /ðə/, never "yee." When you see ye, yt, or y⊤ in a manuscript, read the, that, this. The letter had dropped out of standard English as a live phoneme by the end of the fourteenth century, but it lingered in these stock spellings for centuries.

Two further habits separate a beginner who is guessing from one who is reading.

The first is the minim problem. A minim is a single short vertical stroke, and the letters i, m, n, and u are all built from them. A run of minims — as in minimum — is genuinely ambiguous on the page, because the strokes are identical and only their grouping distinguishes the letters. There is no visual trick that resolves this. You resolve it by knowing what word is plausible in context. This is why fluent reading depends on expectation as much as eyesight.

The second is spelling variation. Standardised spelling is a recent convention. Before the mid-eighteenth century, one writer could spell a name four ways in a single document — Ann, Anne, An, Anne — with no error implied. For genealogists this is the crucial adjustment: match names phonetically, by how they sound, rather than hunting for an exact string. A parish clerk who wrote Smyth one year and Smith the next was not confused. He was writing before the rule you are applying existed.

Putting it together, the sequence a beginner should follow on any new document:

Date and place the hand. Pure secretary, or a mixed secretary-italic hand? Roughly what decade? This sets your expectations.

Read the easy words first. Transcribe what you can read on sight and leave gaps for the rest. The known words anchor the unknown ones.

Attack the troublesome letterforms. Bring the e, c, r, h, and long s to the words you left blank.

Expand the abbreviations, marking the letters you have supplied so your transcription stays honest about what was on the page and what you added.

Resolve the minims and the spelling by context, using the surrounding text and your knowledge of the likely names and formulae.

Check against a model transcription where the tutorial provides one.

Human paleography remains the standard for this work, and the free tutorials from the National Archives, Cambridge, and the Society of Genealogists are the best places to build the skill. There is no substitute for reading pages.

Where machine transcription fits — and where it misleads

At some point the volume of material outgrows what you can transcribe by hand, and the question of software arrives. Here the beginner needs one clear warning and one honest picture of the options.

The warning concerns general-purpose chatbots. It is tempting to paste a manuscript image into ChatGPT, Claude, or Gemini and accept what comes back. Do not treat that output as ground truth. A University of Virginia Library report found that general-purpose models produce fluent, plausible-looking transcriptions that nonetheless contain errors a beginner cannot catch — and, crucially, they offer no confidence flag when they meet a letterform they cannot read. This is the specific danger: the output reads correctly. A model that silently "corrects" Smyth to Smith, resolves an ambiguous minim to the modern-sounding word, or expands an abbreviation the wrong way hands you a clean paragraph with the errors buried where you will never spot them. For genealogy, where a single misread name derails a line for months, a plausible fabrication is worse than an obvious gap.

Dedicated handwritten text recognition (HTR) is the better path. Tools built for historical hands transcribe the image rather than guess at prose. Transkribus, the established platform, publishes a model — Egerton: English Secretary Hand — that reports a 2.89% character error rate on its own validation set; a useful figure, though one earned on a curated test set, so performance on your particular parish register or will is likely to be worse. Transkribus also asks for a subscription and, for the most part, expects users to train or fine-tune models to get the best results on a given corpus.

This is where Leo fits, and it is worth being specific about why. Leo's transcription model, ATR-1, reads Latin-script manuscripts — English, German, Dutch, and the other languages that alphabet records — out of the box, with no per-corpus model training. On a randomized 97-image sample of early-modern English manuscripts from the Folger Shakespeare Library, ATR-1 scored roughly 5% character error rate at its release — 61% fewer errors than the next-best model tested, with Transkribus's Text Titan I at about 13% and the general LLMs far higher (full benchmark here). More to the point for this guide: it is built to preserve what is on the page rather than smooth it into modern prose. The long s stays a long s; the macron is not silently resolved; the archaic spelling survives. That source integrity is exactly the property a beginner needs, because it keeps the machine's errors on the surface — recoverable slips at the level of characters and words — rather than buried in fluent, confident paragraphs. If you want to understand the mechanism behind that distinction, the difference between OCR and HTR for historical documents is the place to start.

Whatever tool you use, the discipline does not change: always correct the machine output against the original image. Software gives you a first pass. Your eye, trained on the letterforms, gives you a transcription you can trust.

The tools will keep improving — and models trained on user corrections do compound in accuracy over time — but none of them removes the need to read. A machine transcription you cannot check is a machine transcription you cannot defend, and for family history a name you cannot defend is a branch you may have to prune later. Learn the secretary e and the long s. Learn to see a macron and supply the missing n. Learn to read ye as the and to match Ann to Anne by sound. Do that, and the documents that stalled your research for months become, page by page, simply things you can read.

How do you read secretary hand as a beginner?

Start by learning the handful of letterforms that differ systematically from modern ones — the secretary e, c, r, h, and the long s — then learn the abbreviations that delete whole letters and syllables, and finally build the habit of reading by context rather than exact spelling. Work from an alphabet chart alongside a real manuscript, transcribing the easy words first and leaving gaps for the rest. Expand abbreviations, resolve ambiguous minims by plausible context, and check yourself against a model transcription. It is a learnable skill built one page at a time, not a gift.

What is the difference between secretary hand and court hand?

Secretary hand is the everyday vernacular script of English-speaking Europe from roughly 1500 to 1700, used for wills, deeds, letters, and parish registers. Court hand is a narrower, more angular script used specifically in English legal records from the thirteenth to eighteenth centuries, descending from a different lineage — a gothic chancery hand. The two overlap in time, and courts used secretary hand too, so you will meet both. If a document looks far more cramped and unfamiliar than the wills and letters around it, you may be looking at court hand, which is a separate study.

No. Secretary hand is overwhelmingly a vernacular script, so the documents you read are in English, German, Dutch, and other spoken languages — not Latin. This is the misconception that trips up newcomers most often. Latin appears only in formulaic patches: legal boilerplate, dating clauses, and standard abbreviations. You do not need Latin to read a Tudor will. What you need is to read early modern English written in an unfamiliar hand, with attention to the letterforms and abbreviations that make the page look harder than the language actually is.

Why does "the" look like "ye" in old handwriting?

Because of the thorn (þ), the Old English letter for the th sound. In many secretary-hand forms the thorn came to look like a y, and early printers — lacking a thorn sort in their type — substituted a y. That is the entire origin of "Ye Olde." The word was always "the," pronounced /ðə/, never "yee." When you see ye, yt, or a similar form in a manuscript, read the, that, this. The thorn had dropped out of standard English as a live sound by the end of the fourteenth century, but it lingered in these stock spellings for centuries.

Can AI transcribe secretary hand accurately?

Dedicated handwritten text recognition can, but general-purpose chatbots like ChatGPT, Claude, or Gemini should not be trusted for it. General models produce fluent, plausible-looking transcriptions that contain errors a beginner cannot catch, and they give no confidence flag when they meet a letterform they cannot read. Purpose-built HTR tools transcribe the image rather than guess at prose. Leo's ATR-1 model scored roughly 5% character error rate on a randomized sample of early-modern English manuscripts — 61% fewer errors than the next-best model tested — and preserves the long s, the macron, and archaic spelling rather than smoothing them away. Whatever tool you use, always correct its output against the original image.

Learn the letterforms, abbreviations, and context-reading habits that make secretary hand legible, from the long s and secretary e to macrons and minims.

How to Read Secretary Hand: A Beginner's Guide

This is a working guide to early modern paleography — the study of how people wrote between roughly 1500 and 1800, and how to read what they left behind. It covers the hands you will meet in a European or North American archive, the marks that trip readers up, the dating conventions that quietly corrupt a citation, and where machine transcription fits into serious research without displacing the historian's judgment.

Early modern paleography is the study of how people wrote between roughly 1500 and 1800, and how to read what they wrote. For a working historian, the practical problem breaks into three parts: identifying the hand on the page (secretary, italic, French bâtarde, German Kurrent, and their relatives), decoding the letterforms and abbreviations that no longer survive in modern writing (the long s, the thorn-derived ye, the macron for an omitted nasal), and choosing a transcription convention — diplomatic, semi-diplomatic, or normalized — and holding to it consistently. Get those three right and the rest is patience.

This is a reference for that work: the hands you will actually meet in a European or North American archive, the marks that trip people up, and the conventions that make a transcription defensible. It sits within the broader craft of reading old handwriting and paleography, and it assumes you would rather understand a page than be told a machine has understood it for you.

Paleography begins with recognition. Before you can read a document you have to know what family of script you are looking at, because the letterforms — and the traps — differ by hand.

Secretary hand is the script most English researchers hit first and hardest. It was the dominant English documentary hand of the sixteenth and seventeenth centuries, and its diagnostic features are unfamiliar enough to stop a modern reader cold: looped ascenders and descenders, a single-compartment a, an e that reads like a reversed e or a c with a tail, a long-tailed r, and h and t built from distinctive crossed strokes. Italic displaced it over the eighteenth century, but it persisted in legal and formal use into the early nineteenth. Cambridge's English Handwriting 1500–1700 resource remains the standard training ground, and if secretary is your immediate obstacle, a dedicated walkthrough of its letterforms is worth the detour.

Italic and humanist cursive arrived from Italy in the late fifteenth century, derived from humanist minuscule, and it is the ancestor of modern English handwriting — which makes it the easiest of the period's hands to read at sight. From it descend the English round hand (copperplate), a seventeenth-century development that became the eighteenth-century standard.

Court hands are a separate difficulty. This family of English legal scripts — Common Law, Chancery, Exchequer — was deliberately conservative and, in the case of Chancery, famously hard; it lingered in formal use into the eighteenth century long after ordinary writing had moved on.

The Continental hands follow their own lineages. French cursive, the bâtarde and notarial hands, dominated French vernacular writing from the sixteenth to eighteenth centuries: dense, heavily looped, and thick with ligatures. The École nationale des chartes maintains Theleme as a reference for it. German Kurrent and its angular chancery variant Kanzleischrift served German vernacular writing from the sixteenth century until 1941; Sütterlin, the simplified teaching form most people have heard of, was only in schools from 1911 to about 1941 and is a late chapter, not the whole story. Spanish procesal — a long-lived notarial cursive — descends from the fifteenth-century cortesana, the ancestor of italic type. Italian mercantesca (the medieval merchant hand) and cancelleresca (the chancery hand from which italic type derives) round out the Romance tradition, alongside Dutch chancery cursive with its own distinctive letterforms.

A note that matters for anyone reaching for a tool later: these are all written in the Latin alphabet, whatever the language. The distinction between script and language is not pedantry — it governs what any recognition system can and cannot do, a point developed in the discussion of which languages and scripts HTR actually reads.

Once you have identified the hand, the individual marks are where transcription errors are made. A handful recur across languages and hands often enough to be worth committing to memory.

The long s (ſ, U+017F). Used medially and initially in English, French, and German from roughly 1500 to 1800 — but not as a final letter, and not before another long s. It looks like an f without the full crossbar, and misreading it as f is the single most common early modern transcription error. It was abandoned in English print around 1790–1810 and persisted in handwriting a little longer.

Thorn and the y-shaped article. Thorn (Þ/þ) had been replaced by th in English by the fourteenth century, but its ghost survives in secretary hand as the y-shaped abbreviations ye, yt, and ys — meaning the, that, and this. They are not "ye olde" anything; the y is a thorn in disguise.

Minims. The short vertical strokes that make up m, n, u, i are the classic ambiguity: a run of them — minimum, say, or a Roman numeral — can be read several ways, and only context resolves it. This is precisely the kind of judgment paleographic competence exists to make.

u/v and i/j. These were variant shapes of single letters — u and v a vowel/consonant pair, i and j likewise — and the modern split only stabilized in the sixteenth and seventeenth centuries. Expect loue for love and iustice for justice.

Abbreviations. Early modern scribes and printers alike carried a large repertoire of space-saving marks. A suspension omits letters at the end of a word; a contraction omits them in the middle, often marked by a stroke through a descender. The macron or titulus (combining macron, U+0304) marks an omitted nasal, so q̃ expands to que and a barred p to per or pro. Standalone signs — brevigraphs — stand for whole words: the Tironian et (⁊, U+204A), the -us/-rum abbreviation, the -em sign. For Latin abbreviations, Cappelli's Dizionario di abbreviature latine ed italiane remains the reference dictionary. When you record these in Unicode, the Medieval Unicode Font Initiative recommendation gives you code points aligned to the standard rather than improvised substitutes.

A correctly transcribed date can still be wrong by a year if you do not know the conventions.

Regnal years count from a monarch's accession, not the calendar year: 14° Eliz. is the fourteenth regnal year of Elizabeth I, which runs across 1571/72. Old Style and New Style is the more dangerous trap. England kept the Julian calendar until September 1752, with the legal year beginning on 25 March (Lady Day), while Catholic Europe had adopted the Gregorian calendar from 1582. A document dated in January, February, or March in seventeenth- or eighteenth-century England therefore often carries what looks to us like the previous year. And in Roman numerals, a terminal j — iiij, vj — is simply a variant final i, not a distinct numeral.

Reading the page is half the task. The other half is deciding how faithfully to represent what you read, and documenting that decision.

Diplomatic transcription preserves everything verbatim: original spelling, abbreviations, punctuation, deletions, insertions. It is the most defensible for scholarly citation and the most laborious.

Semi-diplomatic expands abbreviations — usually with the expansion marked, as yo[u]r — but keeps original spelling and layout.

Normalized transcription modernizes spelling, punctuation, and layout. Legitimate, but only if the convention is stated so a reader knows what has been changed.

The recurring error here is treating archaic spelling as a mistake. It is not. Publick, shew, yeomanry — these are orthographically correct for their date, because spelling was not standardized. Silently modernizing them is itself a transcription error, and one that quietly destroys evidence about scribe, region, and date. When you formalize any of this for publication, TEI P5 chapter 12 provides the standard encoding for primary sources, and chapter 11 for manuscript description.

Where machine transcription fits — and where it does not

At some point the volume of material outruns the hours. A research project resting on hundreds of pages of French notarial cursive or English secretary hand cannot be transcribed by eye alone on any realistic timeline, and this is where automated transcription enters the workflow. It is worth being precise about what it does and does not change, because three misconceptions do real damage.

Stock modern OCR — Tesseract, ABBYY, Google Vision, AWS Textract — was trained on modern typefaces and fails systematically on early modern print: it reads the long s as f, drops ligatures, and silently loses the macron. The Texas A&M eMOP project documented baseline character error rates above 10% on early modern print, with dedicated post-correction pipelines needed to push that lower across some 307,000 pages. General OCR is engineered for a page that does not look like your page.

Multimodal models can transcribe a historical page with no training, and their headline numbers are not embarrassing. But the failure mode is the problem. General LLMs err by silent normalization — modernizing spelling, regularizing capitals and punctuation — and by producing plausible readings with no statistical grounding in the image. On genuinely difficult English hands, a recent preprint (Humphries, Leddy & Downton, 2024) found general models roughly comparable to a stock Transkribus model and closer still after self-review, but that result rests on a single corpus of eighteenth- and nineteenth-century English hands, has not completed peer review, and — like every method here — still requires a paleographically literate reader checking the output against the page. A character-level error from a specialist tool is visible and recoverable. A fluent fabrication is neither, which is exactly why it is dangerous.

It does not. Fine-tuned specialist models reach low single-digit character error rates only on hands inside their training data; on unseen hands they run far higher, and the training burden is around 75 ground-truth pages per hand. You still need paleographic competence to check the output, resolve the minims, and decide which readings to trust. The mechanics of that pipeline — segmentation, recognition, why accuracy tracks how well the model matches your hand — are worth understanding, and are set out in how HTR actually works and in the distinction between OCR and HTR.

Source integrity as the governing principle

The principle that should govern the choice of tool is the one this whole guide turns on: source integrity. A transcription is evidence only if it preserves what is on the page. This is where Leo's transcription model, ATR-1, takes its position — it is built to transcribe what is written rather than to smooth it into modern, plausible prose. The long s stays a long s; the macron is not silently resolved to the letters it stands for; the archaic spelling and the strikethrough survive into the transcript. It reads Latin-script manuscript and printed material across English, French, German, Spanish, Italian, Dutch, and Latin among other languages, out of the box, with no per-hand model to train first. On a randomized 97-image sample of early modern English manuscripts from the Folger Shakespeare Library, ATR-1 recorded roughly a 5% character error rate at release — the lowest in a comparison that included Transkribus's Text Titan I (~13%) and several general LLMs; the full benchmark data is published here. None of that removes the reader from the loop. It moves the bottleneck from deciphering every character to verifying a draft — which, when the alternative is transcribing a diary of 900 pages by hand, is the difference that lets the analysis begin.

Tools change the economics of transcription; they do not change the discipline. Knowing that ye is a thorn, that a January date in 1650s England may mean 1651 to us, that publick is not a typo — this is knowledge that lets you catch the machine's errors, defend your readings, and understand what a scribe was actually doing on the page. Paleography was never only about deciphering marks. It is about reading a document closely enough to know what it is. That competence is what makes a transcription citable, and it stays with the historian, whatever reads the first draft.

Early modern paleography is the study of how people wrote between roughly 1500 and 1800, and how to read what they left behind. For a working historian the task breaks into three parts: identifying the hand on the page (secretary, italic, French bâtarde, German Kurrent, and their relatives), decoding the letterforms and abbreviations that no longer survive in modern writing (the long s, the thorn-derived ye, the macron for an omitted nasal), and choosing a transcription convention — diplomatic, semi-diplomatic, or normalized — and holding to it consistently. Get those three right and the rest is patience.

Why is the long s so often misread in old documents?

The long s (ſ) is the single most common early modern transcription error because it looks like an f without the full crossbar and is routinely misread as one. It was used medially and initially in English, French, and German from roughly 1500 to 1800 — but never as a final letter and not before another long s. English print abandoned it around 1790–1810, and handwriting held onto it a little longer. Recognizing it on sight is one of the small competences that separates a defensible transcription from a corrupted one.

Yes — the y in forms like ye, yt, and ys is a thorn in disguise, standing for the, that, and this. Thorn (Þ/þ) had been replaced by th in English by the fourteenth century, but its ghost survived in secretary hand as these y-shaped abbreviations. So "ye olde" was never pronounced with a y sound; it is simply an old spelling of "the." Reading it as a literal y is one of the recurring traps in early modern documents, and knowing the history resolves it immediately.

Can OCR or an LLM transcribe early modern documents accurately?

Not reliably on their own. Stock modern OCR — Tesseract, ABBYY, Google Vision, AWS Textract — was trained on modern typefaces and fails systematically on early modern print, reading the long s as f, dropping ligatures, and losing the macron; the eMOP project documented baseline character error rates above 10% on early modern print. General LLMs err by silent normalization, modernizing spelling and punctuation and producing plausible but ungrounded readings. Both still require a paleographically literate reader checking the output against the page, because a fluent fabrication is far harder to catch than a visible character-level error.

Why does a January date in 1650s England sometimes mean the following year?

Because England kept the Julian calendar until September 1752, with the legal year beginning on 25 March (Lady Day) rather than 1 January. A document dated in January, February, or March in seventeenth- or eighteenth-century England therefore often carries what looks to us like the previous year — an Old Style date that must be adjusted to New Style. Catholic Europe had adopted the Gregorian calendar from 1582, so the two systems ran in parallel for well over a century. Ignoring the convention can put a correctly transcribed date wrong by a full year.

Covers the hands, letterforms, abbreviations, and dating traps in documents from 1500 to 1800, with transcription conventions and where machine tools fit in.

Early Modern Paleography: A Working Guide

Reading Early Modern Paleography: A Working Guide to the Hands, Letterforms, and Transcription

Reading Old Handwriting & Paleography

Manuscript Abbreviations and Ligatures: A Reader's Guide to Scribal Shorthand

How to Read Secretary Hand: A Beginner's Guide

Reading Early Modern Paleography: A Working Guide to the Hands, Letterforms, and Transcription