Can OCR Read Cursive? Joined Letters and Their Limits

OCR can read cursive sometimes. Software built or trained for handwriting can return usable text from even, well-spaced cursive, while an engine made for printed type will not. What decides it is the software, how regular the writer's hand is, and the scan. No tool reads every cursive page, so test your own.

The rest of this post is about the cursive part of the problem only: letters that join. It explains why joining matters to software, which kinds of cursive turn up in a legal file, and how to sort a stack of cursive pages so that your time goes to the ones worth it.

Can OCR read cursive handwriting?

OCR can read some cursive handwriting. Three things decide the result: whether the software was made for handwriting at all, how consistent and separate the writer's words and lines are, and how well the scan kept the pen strokes. Change any one and the same tool gives a different answer.

The software

Two vendors' own pages show that the kind of software matters. Transkribus, a platform for transcribing handwritten documents, tells users in its help page on choosing a text recognition model that "There does not exist a general model for all the handwritings". reMarkable's support article on converting handwritten notes into text says that on its tablet "Block letters and cursive handwriting are recognized". Those are statements about what each product accepts. Neither says how your client's writing will come out.

The hand

A cursive page is easier for software when the same letter looks alike each time it appears, words have clear gaps between them, lines do not touch, and loops are open. It is harder when the writer changes slant, runs words together, or reduces the ends of words to a wavy line.

The scan

Cursive depends on thin connecting strokes. A black-and-white scan at a low resolution can drop those strokes or fill in the loops, and a letter "e" with a closed loop looks like an "i" or a "c". Scan cursive in grayscale or color so the faint joins survive.

Schema Law reads uploaded pages into text and tables. Its homepage says recognition is strongest on clean print and weaker on handwriting, and that nothing marks a passage as uncertain. Treat any text it returns from a cursive page as a draft to check against the image. The free plan on the pricing page allows one document a day of up to three pages, which is enough to try a sample.

Why are joined letters harder for software than printed capitals?

Joined letters are harder because the software has to work out where each letter starts and stops before, or while, it decides what the letter is. Printed capitals arrive already separated by white space. Cursive arrives as one continuous line per word, and the cut points are a matter of judgment.

Write the word "minimum" in cursive and look at it. It is fifteen near-identical downstrokes joined at the bottom. Only the count of strokes and the dots tell a reader which are "m", "n", "i" and "u". A person reads it because they know the word. Software that cuts the line into single characters first has no safe place to cut.

Three features of cursive add to the difficulty:

• Joins change the letter. A cursive "o" followed by "n" ends in a high connecting stroke. Followed by "a", it ends differently. The same letter has several shapes depending on its neighbours.
• Parts of letters are written late. The writer finishes the word and then goes back to cross the "t" and dot the "i". Those marks can land over the wrong letter.
• Loops overlap. The tail of a "g" on one line can run through the top of an "l" on the line beneath, so even separating the lines is uncertain.

One answer in software designed for handwriting is to read a whole line as a sequence instead of cutting out characters. Transkribus's page on automatic layout recognition describes finding a baseline that runs along the bottom of each handwritten line and calls it "the most important reference point for text recognition". The post on handwritten text recognition explains that approach. Joining is one of several reasons script defeats software, and the wider list is in the section on why handwriting is harder than print.

Cursive styles a legal file contains: school script, signatures and scrawl

Three kinds of cursive show up in legal files, and they behave differently. School script is written to be read by someone else. A signature is written to be recognized, not read. Scrawl is written at speed for the writer alone. Hope for the most from the first and do not count on a name from the second.

A signature is the clearest example of writing that is cursive in form and not meant to be read letter by letter. Its job is to look like the same mark each time. If you need to know who signed, look for a typed or printed name, a notary's certificate, or the body of the document.

When the cursive mark is a signature, compare it with the printed name, and still read the date, any amount, and the full name in the body. Give a handwritten letter the same light pass. Keep the image as documentary evidence. Note who scanned it so you can state the authentication and a short chain of custody. Quote that image when the best evidence rule applies.

Scrawl has a second problem besides the joining. The writer leaves out letters and whole words, and uses private shorthand that only makes sense to someone who knows the subject. Software can only read what the pen put on the page.

Handwriting from earlier centuries is a separate case again. The letters themselves have different shapes, so a modern reader has to learn them before the question of software even arises. The post on how to read old handwriting covers those hands.

Does asking a writer to print make a page readable by software?

Printing helps, and it does not guarantee a readable page. Separate capital letters remove the joining problem. They do not remove faint ink, crowding, letters that touch, or a writer who slips back into script halfway down the page. Ask for printing where you control the form, then test what comes back.

Forms are designed around this. The boxes, the "use block capitals" instruction and the request for black ink exist to get one clear character at a time, and the section on hand-printed boxes versus cursive explains the kind of recognition built for that.

If your office sends out intake forms or questionnaires that clients fill in by hand, the wording, the boxes and the space on the page are yours to set. The post on handwritten forms has a checklist for designing a form so handwritten answers convert.

You cannot ask for any of this after the fact. A letter a decedent wrote in 1998, or the opposing party's handwritten notes, exists in the hand it was written in. Do not ask a witness to recopy their own notes in print so that software can read them. The copy is a new document, and the original is still the one that matters. Scan the original as it is.

Sorting cursive pages by how much a reader can make out

Sort cursive pages by reading one line of each. How easily a stranger to the handwriting reads that line tells you how much checking any typed version will need, and who should produce it. What software makes of the same page you learn only from a sample, so run one before you commit a whole file to one route.

The one-line test: pick a line from the middle of the page, not the heading, and read it aloud at normal speed without the context of the lines around it.

A short example. A probate file for the estate of Harold Brennan contains forty pages of his letters to a niece and a pocket notebook of figures. One line from a letter reads straight through, so the letters go in pile A. One line from the notebook yields "paid" and a number, so the notebook goes in pile C and a paralegal types the entries that matter.

Whichever pile a page lands in, keep the scan with the text. A typed version of cursive is someone's reading of it, whether the someone is software or a person, and the image is what settles a disagreement.