# OCR Old Documents: Long S, Ligatures and Worn Type

- Author: Schema Law (https://www.schemalaw.com/)
- Published: 2026-04-06T05:45:22Z
- Updated: 2026-10-11T19:22:20Z
- Category: Difficult scans
- Canonical: https://www.schemalaw.com/blog/ocr-old-documents

> OCR can read old documents printed before the typewriter, with predictable errors. See what the long s, ligatures, blackletter and worn type do to the text.

OCR can read old documents set in roman type, but early printing leaves a recognizable trail of errors. The long s comes out as an f, joined letter pairs split or vanish, worn type breaks into fragments, and spelling correction rewrites words that were right. Blackletter pages need a tool built for them, or a typist.

## Can OCR read old printed documents?

Yes, with limits that depend on the typeface and the state of the page. A cleanly printed page in roman type converts into text you can search, and you then repair a known set of errors. Blackletter, heavy foxing and broken type each push the result toward something you proofread line by line.

"Old" here means printed with movable type on a press, before the typewriter. Typed pages have their own faults, set out in the guide to [OCR on typewritten documents](https://www.schemalaw.com/blog/ocr-typewritten-documents). Pen-written registers and minutes are a different job again, and the section on [old court records, cursive and faded ink](https://www.schemalaw.com/blog/ocr-handwritten-will#old-court-records-cursive-and-faded-ink) covers them.

### Grade the page before you convert it

Look at one page at 200% zoom and sort it into a row of this table. It tells you how much checking to plan for.

| What the page looks like | What to expect from the text | Where the errors gather |
| --- | --- | --- |
| Roman type, even ink, pale paper | Searchable text with a short list of repeated mistakes | Long s, joined letters, old spellings |
| Roman type with italics for names and Latin | The upright text reads; the italic runs read worse | Proper names, case names, Latin phrases |
| Roman type, worn or under-inked | Words split in two, letters dropped | Thin strokes: l, i, t, f, long s, punctuation |
| Roman body with a blackletter heading | Body reads; the heading is garbled or missing | The first lines of each instrument or chapter |
| Blackletter or Fraktur throughout | Unreliable from a general-purpose tool | Everything |
| Printed form completed by pen | The printed words read; the pen entries are a separate question | Names, dates, sums, descriptions |

### Run a three-page test

Pick the cleanest page, an ordinary page and the worst page in the volume. Convert all three and read each result beside its image. Schema Law's free plan takes one document a day of up to three pages, which fits this test exactly, and you see the parsed pages on screen before you export anything. The site says recognition is strongest on clean print, and nothing marks a passage as uncertain, so the reading is yours to do. The [steps from upload to export](https://www.schemalaw.com/#how-it-works) are on the home page.

If the volume is still bound, the capture comes first. The guide to [converting a scanned book to text](https://www.schemalaw.com/blog/how-to-convert-scanned-book-to-text) deals with gutters, spreads and page order.

## What is the long s, and why does OCR read it as an f?

The long s is an old form of the lowercase s, drawn as a tall stroke with a hooked top. It looks like an f without a full crossbar. Software trained on modern print has a slot for f and none for this shape, so it returns the letter the shape most resembles.

The National Archives, in its [transcription tips for volunteers](https://www.archives.gov/citizen-archivist/transcribe/tips), puts it this way: "This letter that looks like a lowercase f is an s." The same page says you will see these "in handwritten and printed documents into the early part of the nineteenth century", and its own example is "houfe" for "house". Its instruction to transcribers is to type an ordinary s.

### What the error looks like in converted text

A sentence from a fictional session law shows the pattern. The page reads, to a modern eye, "the said trustees shall cause the same to be sold, and possession delivered". The converted text may come back like this:

```text
the faid truftees fhall caufe the fame to be fold, and poffeffion delivered
```

Six words are wrong, and two of the wrong words are real English. "Fame" and "fold" pass any spelling check. That is the part that matters for search. If you look for "sold" in the converted volume, this sentence does not appear.

### Words where f and s both make sense

Some pairs cannot be settled by a word list, because both readings are words. Read these against the image wherever the meaning turns on them:

| Converted text shows | The page may say |
| --- | --- |
| fame | same |
| fold | sold |
| fee | see |
| fight | sight |
| fuit | suit |
| fate | sate, or a real "fate" |
| fought | sought |
| feal | seal |

To tell the two letters apart on the image, look at the crossbar. On the pages in front of you, check whether the f has a bar that crosses the stem to the right, and whether the long s has only a small nub on the left or none. Confirm it on a word you are sure of, such as "of" or "said", and then apply what you see to the doubtful word.

### Repairing it without making it worse

Do not replace every f with s. Work in this order:

1. Keep the untouched export as its own file.
2. In a copy, search for letter pairs that do not begin an English word, such as "fh" and "ft", and for the misread forms of frequent words: "faid", "fuch", "thofe", "thefe", "alfo", "muft", "poffeffion". Take care with "firft": the page word is "first", and only the second f is wrong.
3. Fix each pattern with a whole-word replacement and read the hits as you go.
4. For the real-word pairs in the table, read each occurrence against the image.
5. When you search the volume later, search both spellings: "sold" and "fold", "seal" and "feal".

## Blackletter, Fraktur and decorative typefaces in old legal printing

Blackletter and Fraktur are heavy, angular typefaces in which many letters are built from the same close vertical strokes. A tool made for modern roman type has little to match them against. Expect a general-purpose converter to fail on a full blackletter page, and plan for a specialized tool or a human typist.

### Why the shapes defeat a general tool

In blackletter, a run of "minim" strokes makes up i, n, m and u, so a word such as "minimum" is a fence of near-identical lines. Capitals are ornate and bear little relation to their roman equivalents. Several letters have forms a modern reader has to learn. None of that is damage. The page can be perfectly printed and still be unreadable to software that has never been shown the alphabet.

Some engines ship separate recognition data for this kind of type. The open-source Tesseract project's [list of data files by version](https://tesseract-ocr.github.io/tessdoc/Data-Files-in-different-versions.html) includes entries described as "German - Fraktur (now deu_latf)", a script file named "Fraktur", "English, Middle (1100-1500)" and "Latin". The existence of a data file says nothing about how well it reads your page. It tells you that the question "which alphabet was this tool taught?" has more than one answer, and that you should ask it of whatever you use. The post on [typeface features that help or hurt recognition](https://www.schemalaw.com/blog/best-font-for-ocr) covers the general point.

### Display type on an otherwise readable page

A page set in roman type can still open with a line in blackletter or an engraved script. Look for it in these places:

- The opening words of a deed or bond, such as "This Indenture" or "Know all Men by these Presents".
- The title line of an act or chapter.
- The name of the court or the company across the head of a printed form, certificate or share.
- A large decorated initial at the start of a chapter, which may come out as a stray character or a picture.

The body converts and the heading does not, so the converted text starts mid-sentence. Type the display line by hand from the image and mark it as typed by you.

### Ligatures, italics and swash letters

A ligature is two or three letters cast as one piece of type: "ct", "st", "ff", "fi", "ffi", and the long s joined to a following letter. The joined shape can come back as one wrong character, as the right letters, or as nothing. Check "ct" in "act" and "section", and the double long s in words such as "assigns" and "possession".

Old italic type can carry swash letters, capitals and tails drawn with a flourish that a tool taught on modern italics has not met. Check what your volume sets in italic. Where it is the proper names, case names and Latin, those are the words a lawyer searches for, so read every italic word against the image.

## Uneven inking, broken letters and foxed paper in letterpress pages

Letterpress printing pressed inked metal type into paper, one sheet at a time. Ink varies across the page, worn type prints with gaps, and old paper has spotted and browned. Each fault gives software either too little letter or too much background, and the fix starts with how the page is scanned.

| What you see on the image | What it does to the text | What to try |
| --- | --- | --- |
| One side of the page pale, the other heavy | Missing words on the pale side, filled letters on the heavy side | Grayscale scan; check the pale margin first |
| Hairline strokes missing from letters | "h" read as "li", "m" as "rn" or "in", "e" as "c" | Higher scan resolution; proofread names |
| Brown spots (foxing) scattered over the text | Stray periods, commas and accents; a spot on a letter changes it | Grayscale or color, never pure black and white |
| Print from the other side showing through | Ghost letters between the lines, mirrored | A sheet of black paper behind the leaf while scanning |
| Dark halo around each letter where ink spread | Letters touching, read as one wrong character | Nothing at capture; proofread |
| Browned edges and a paler center | Text near the edges lost in a black and white scan | Grayscale; do not crop tight |

### Scan mode decides more than the software does

A black and white scan turns every pixel into ink or paper. On a spotted, browned page that choice goes wrong in both directions: foxing becomes ink, and pale strokes become paper. Grayscale keeps the difference between a brown spot and a black letter. The comparison of [grayscale and black and white scanning](https://www.schemalaw.com/blog/grayscale-vs-black-and-white-scanning) explains the mechanism, and the post on [the best DPI for OCR](https://www.schemalaw.com/blog/best-dpi-for-ocr#why-small-print-needs-more-dots-per-inch) covers the small type of footnotes and sidenotes.

If you received the scan and cannot redo it, sort the pages. Convert everything, then proofread in full only the pages you will quote or cite.

### Damage is a separate problem

Tide lines, tears, tape and insect holes belong to the paper's later life, not to the printing. The guide to [OCR on a stained document](https://www.schemalaw.com/blog/ocr-coffee-stained-document) covers physical damage, including what to do when the damaged sheet is itself evidence.

## Why does automatic spelling correction make old text worse?

Correction assumes that an unfamiliar word is a mistake and swaps in the nearest familiar one. In an old document the unfamiliar word is what the printer set: an older spelling, an abbreviation, a surname or a Latin term. The correction replaces a faithful reading with a modern word the page does not contain.

The swap can happen in three places, and you control only two of them:

- **Inside the recognition step.** Some tools weigh candidate readings against a word list. The post on [whether OCR uses a dictionary](https://www.schemalaw.com/blog/does-ocr-use-a-dictionary) explains how that changes output.
- **In your word processor.** Autocorrect rewrites words as you paste or type. Turn it off for the transcript file before you paste.
- **In your own find and replace.** A replace-all that fixes one long s error can break a different word.

### What gets rewritten

| On the page | A correction may give | Why it matters |
| --- | --- | --- |
| shew, shewn | show, shown, or "shrew" | A quotation must match the source |
| gaol | goal | A different word |
| publick, compleat, chuse | public, complete, chose | "Chose" has its own legal meaning |
| sd, afsd, wch, yr | Nothing sensible | Shortened forms a reader takes as said, aforesaid, which, your |
| ss. | "ss" or "as" | Commonly read as a venue mark in a caption; confirm in your document |
| habendum, scilicet, nisi | A nearby English word | Latin terms that carry meaning in the instrument |
| A surname such as Boteler | Butler | The person in the chain of title |

The National Archives' transcription guidance takes the conservative line. It asks volunteers to "transcribe the word as spelled or abbreviated in the document", and tells them not to correct it. For a legal file the reason is practical: the converted text has to find the page and support an exact quotation.

### Keep two versions

Keep a literal file and a working file. The literal file follows the page, with only the long s typed as s. The working file is a copy in which you may modernize spelling for reading and searching, with a line at the top saying that you did. Quote from the literal file and check the quotation against the image. Old numerals and sums need the same care; the post on [why OCR gets numbers wrong](https://www.schemalaw.com/blog/why-does-ocr-get-numbers-wrong) lists the swaps to look for.

## Session laws, early reports and recorded instruments a lawyer still cites

Three kinds of early printing still reach a working file: session laws and other statute volumes, early volumes of case reports, and printed instruments from a chain of title. Each carries page furniture that modern printing dropped, and each has its own fields worth checking by eye.

### Session laws and statute volumes

Check these on a converted page of statutes:

- **Chapter and section numbers.** Roman numerals in worn type lose strokes. "XLVIII" can come back as "XLVII" or "XLVIIl".
- **Sidenotes.** Summaries printed in the margin beside the text may be woven into the body lines. The post on [reading order in multi-column pages](https://www.schemalaw.com/blog/can-ocr-handle-two-column-documents) explains why.
- **Catchwords.** If the last line of the page holds a single word at the right, and that word repeats as the first word overleaf, it is a catchword. Converted text carries it twice.
- **Dates of passage.** Read the year against the image.

### Early reports

In an old volume of reports, check how the name of the case, the term of court and the names of counsel are set. Italics and small capitals are where a conversion is weakest. Look for a page number printed inside the text or in the margin, which can mark the paging of an earlier edition. If your citation depends on a page number, take it from the image and not from the converted text.

### Printed instruments and recorded forms

A deed, bond, patent or certificate may be a printed form with blanks completed by pen. The printed words convert. The pen entries are a separate reading job. Schema Law's site says recognition is weaker on handwriting, and the same caution applies to any tool. Type the names, dates, sums and land description from the image. For the hands themselves, see the guide on [how to read old handwriting](https://www.schemalaw.com/blog/how-to-read-old-handwriting).

A bargain and sale deed from that era may also carry a restrictive covenant in smaller type under the granting clause. If the land is in a boundary dispute, you check every distance the worn type broke, and you look for an encroachment noted in the margin in a later hand. Type those notes yourself.

### A worked example

Dana Okafor, a paralegal, is asked for the text of a private act said to vest a ferry right in a fictional company, the Marrow Creek Bridge Proprietors. She has a scan of the volume from a library. Her steps:

1. She converts the three pages of the act and the title page of the volume.
2. She searches the text for "ferry" and finds it. She searches for "sold" and finds nothing, then searches "fold" and finds the clause.
3. She saves the export unchanged, makes a copy, and repairs the long s patterns in the copy.
4. She types the blackletter title line by hand and marks it in brackets as typed from the image.
5. She reads the chapter number, the date of passage and every proper name against the scan.
6. She exports a searchable PDF of the pages for the file, so the page image travels with the text. The [export formats](https://www.schemalaw.com/#formats) list shows that option beside Word and plain text.

The memo quotes the act from her literal file and cites the page from the image.

## Try it on a real document

Upload a scanned PDF and export it as Word, a searchable PDF or plain text on the free plan. Your documents are never used to train an AI model. [Start free](https://www.schemalaw.com/api/auth/google)