Best Handwriting to Text Converter: How to Test One
By Schema LawPublished Updated
No single product is the best handwriting to text converter, because results depend on the writer, the pen and the scan. The best converter for your firm is the one that makes the fewest errors that matter on your own handwritten pages. You find it by running the same small test set through each candidate.
Is there one best handwriting to text converter for every hand?
No. A handwriting to text converter that reads one person's writing well can fail on the next person's, and no ranking can tell you which side your documents fall on. Handwriting varies far more than print does, and every tool has seen some styles of writing more than others.
Four things make a general ranking unreliable for handwriting in particular.
The writer is the biggest variable. With printed documents, two firms' files look much alike, so one firm's experience transfers to another. With handwriting, your files are the handwriting of specific people: one treating physician, three adjusters, a client who writes in capitals. A converter's result on a reviewer's sample notes says little about them.
Styles differ in kind. Careful block print, joined cursive, a mix of the two and private shorthand are close to four separate problems. A tool can be strong on the first and useless on the third.
The image decides a great deal. The same page scanned in grayscale and in harsh black and white can give two very different results from one tool. A comparison that does not hold the image constant is comparing scanners.
Tools change. Recognition software is revised from one version to the next. A review from two years ago describes a product that may no longer exist in that form, for better or worse.
So when someone asks what the best OCR for handwriting is, the useful reply is a method. The rest of this post is that method: know the kinds of tool on offer, build a small test set from your own files, score it in a way that reflects legal work, and ask the vendor a short list of questions before money changes hands.
One expectation to set first. Even the tool that wins your test will produce a draft. A person still has to read it against the page before anyone relies on it. The test tells you which draft costs the least to correct.
What kinds of handwriting converter can a firm choose from?
A firm can choose among six kinds of handwriting converter: note apps that convert stylus writing, phone scanning apps, the OCR built into PDF software, document parsing services, platforms trained on one writer's hand, and people who type. They solve different problems, so the first step is to match the kind to your documents.
| Kind | What it was built for | Fits a firm when | Watch for |
|---|---|---|---|
| Note-taking apps with stylus conversion | Turning what you write on a tablet into type, using the recorded pen movements | Lawyers want their own tablet notes typed | It works from live pen strokes; a scan of paper is a different input and may not be supported at all |
| Phone scanning apps | Capturing a page with a camera and reading the text in it | One or two pages on the road | Aggressive image cleanup that erases faint ink; consumer terms of service |
| OCR inside PDF and office software | Adding a text layer to printed documents | Typed files with an occasional handwritten entry | Often designed around printed type; read what the documentation says about handwriting |
| Document parsing and recognition services | Reading mixed pages, print and handwriting, into text and tables | Scanned files where handwriting sits among typed pages | Wrong words that read as plausible; confidentiality terms |
| Platforms trained on a specific hand | Reading large sets by one writer or period after training on samples | A long diary, ledger series or historical record set | Setup effort: someone must transcribe sample pages first |
| Human transcription | A typed transcript made by a person | The words will be quoted, or software fails the test | Cost, turnaround and who sees the document |
Two distinctions in the table save the most wasted effort.
The first is between writing captured as you write and writing captured as a picture. A tablet records the order and direction of each stroke, which is strong evidence about the letter. A scan has only the finished marks. Handwriting to text software built for the first job is not automatically any good at the second, so a lawyer's happy experience with a note app does not carry over to a box of paper.
The second is between older print-first engines and newer systems trained on whole pages. The post on AI OCR explains what model-based systems do differently, including the trade-off that matters here: a system that uses context to read a faint word can also use context to supply a wrong one.
A product can belong to more than one row. What matters is which row your documents need, and whether the product's own documentation claims that use.
Building a test set from your own handwritten pages
Build the test set from eight to twelve pages written by the people whose handwriting your firm actually receives, with a typed answer for each page prepared by a person. Organize it by writer, not by document type, because the writer is what the converter succeeds or fails on.
Choose the writers
List the hands that recur in your work. For an injury practice that might be two treating providers, an adjuster and a typical client. For an estates practice it might be elderly clients' letters and an old attorney's file notes. Pick three to five writers and take two pages from each.
Cover the spread on purpose:
- at least one writer who prints and one who writes cursive
- one page in pencil or light ink, if your files contain them
- one page dense with names, dates and figures
- one printed form with handwritten entries
- one writer you consider hard to read
Leave out the single worst page in the office. A page no person can read tests nothing. Leave out typed pages too, unless you are also choosing a tool for print, in which case the general method for running your own test on real documents is the place to start.
If those pages come from an estate file, include a letter that names an heir and describes a bequest or a devise. Score a wrong heir the same way you score a wrong date. A line about undue influence is one of the items that would hurt if the draft invents a word the writer never put down.
Settle what you may upload
A trial is still a disclosure of client information to a vendor. Use pages your firm's policy allows: public filings, closed-matter material cleared for the purpose, or pages a colleague writes out for the test in their usual hand. Read each vendor's terms on retention and training before the first upload.
Make one image per page and reuse it
Scan each page once, with care, and send that identical file to every candidate. If you rescan between tools, you have changed the test. The post on scanning handwriting into text covers the settings that keep faint strokes in the image.
Type the answers
For every page, have a person type what the page says, line for line, and have a second person check it. This typed version is what you score against. Where the humans cannot read a word, mark it as unreadable in the answer and exclude it from scoring. A converter should not be penalized, or rewarded, for a word nobody can verify.
Then underline, in each answer, the items that would hurt if wrong: every name, date, amount, dosage, address, and every "not", "no" and "never". Those are your critical items.
Freeze it
Put the images, the typed answers and a one-page description in a folder and do not edit them. The same set will judge next year's version of the winning tool and any new candidate that appears.
How do you score a handwriting converter's output?
Score a handwriting converter by counting, for each page, the critical items it got wrong, the lines it dropped, the words it invented and the minutes a person needs to correct the draft. A single accuracy percentage hides the difference between a misspelled "the" and a wrong date.
Vendors and researchers often report a character or word error rate, which is the share of characters or words that differ from a correct transcript. The post on how OCR accuracy is measured explains those rates. They are fair summaries, but two tools with the same rate can be very different to work with, so count by kind.
| Error kind | What it looks like | Why it gets its own count |
|---|---|---|
| Critical item wrong | "Hartney" for "Hartley"; "2019" for "2014"; a missing "not" | One of these can change advice or a filing |
| Dropped text | A line, a margin note or a postscript absent from the output | Nothing in the draft shows that something is missing |
| Invented text | Words in the output that are not on the page | The draft says something the writer did not |
| Ordinary word wrong | "recieve" for "receive"; "form" for "from" in a harmless sentence | Costs correction time, rarely changes meaning |
| Order wrong | Lines or columns out of sequence; a margin note spliced into a sentence | Makes correct words misleading |
A score sheet
Use one row per page and per tool.
| Writer and page | Critical items on page | Critical wrong | Lines dropped | Invented passages | Minutes to correct |
|---|---|---|---|---|---|
| Provider 1, p. 1 | 14 | ||||
| Provider 1, p. 2 | 9 | ||||
| Adjuster, p. 1 | 17 |
To fill in the last column, give the draft and the image to the person who would do the work in practice, and time them correcting it to match the typed answer. That number is the closest thing to the real cost of the tool.
Reading the results
Here is how a filled-in sheet might read, with invented figures. Across ten pages there are 120 critical items. Candidate one gets 11 wrong, drops no lines and takes 62 minutes to correct in total. Candidate two gets 8 wrong but drops four lines on the two cursive pages and takes 95 minutes, because the checker had to find the gaps. On raw critical errors candidate two looks better. For a firm that mostly needs searchable, correctable drafts, candidate one is the safer choice, since its failures are visible and its drafts are quicker to fix. A firm whose handwriting is nearly all cursive might read the same sheet the other way and decide neither tool is good enough for that pile.
Look at the results by writer before you total them. A tool that is near-perfect on three writers and hopeless on a fourth is a good tool with a known exclusion. An overall average would hide that.
Finally, compare each tool with no tool. Time a person typing one of the test pages from scratch. If correcting the best draft is not clearly faster, the best handwriting recognition software for that writer is a typist.
Handwriting questions to put to a vendor before you pay
Ask a vendor whether the product was built to read handwriting, how it behaves when it cannot read a word, what happens to the pages you upload, and whether you can test your own documents first. Ask for the answers in writing, or find them in the published documentation.
About the reading itself
- Does your documentation say the product reads handwriting, or only printed text? Where?
- Does that cover cursive, or only printed-style writing?
- Which languages does handwriting recognition cover?
- When the software cannot read a word, what does it do: leave a gap, output its best guess, or mark the word in some way?
- Does the output show which words the software was unsure of? The post on OCR confidence scores explains what such a signal can and cannot tell you.
- On a typed page with handwritten notes, where do the notes appear in the output?
- Is the same recognition used in every plan and every export format?
About the documents
- Are uploaded pages used to train or improve any model? The post on OCR that does not train on your data explains how to read the answer.
- How long are uploads kept, and can we delete them ourselves?
- Who at the vendor, or at its suppliers, can see a page?
About the purchase
- Can we run our own test pages before paying, and how many?
- What are the limits on file size and pages per document?
- How will we learn that the recognition has changed, so that we can rerun our test?
Be wary of two kinds of reply. One is an accuracy percentage with no description of the handwriting it was measured on. The other is a demonstration on the vendor's sample page. Neither says anything about your writers.
To show how published terms answer some of these, take Schema Law's. Its homepage says recognition is strongest on clean print and weaker on handwriting, and that nothing marks a passage as uncertain, which answers questions 4 and 5 without a call. Every plan lists the lines "Your documents are never used to train an AI model" and "Delete any document, at any time". Nothing is deleted automatically, so a test upload stays in the account until you remove it. Uploads are a PDF, PNG or JPEG of up to 20 MB, and the free plan takes one document a day of up to three pages, so a test set goes through a few pages at a time. The pricing page lists the limits on the paid plans. None of that tells you how it will read your adjuster's notes. Your score sheet does.
When a transcription service beats any handwriting software
A transcription service, or a careful typist in your own office, beats any handwriting software when the exact words will be quoted or filed, when the test shows correction takes as long as typing, or when the writing is shorthand that needs a reader who knows the subject.
The cases are easy to recognize once you have test results. The post on handwritten notes sorts work by purpose in its section on when a person still has to transcribe it. The list here starts from your score sheet.
- The winning tool still loses to typing. If the best draft for a writer took longer to correct than a fresh transcript took to type, stop testing and type.
- The document is short and important. A two-page handwritten statement at the center of a dispute needs a verified transcript whatever software says. The draft saves little, and the checking is the real work.
- The handwriting is specialized. Clinical abbreviations, trade shorthand and old script need a reader with the vocabulary. A medical transcriptionist or a person trained in older hands will resolve words that software and a general typist both miss.
- Errors must be visible. A person can write "[illegible]" where they are stuck. Software often writes a confident guess.
If you send work outside the firm, treat the service as you would any vendor that handles client papers, and ask a short set of questions.
| Ask the service | Why |
|---|---|
| Who types the pages, and where are they? | You are disclosing client documents to those people |
| What confidentiality terms do typists work under? | Your duties follow the document |
| Is each transcript checked by a second person? | A single reader's errors are invisible |
| How are unreadable and uncertain words marked? | You need the doubts shown, in a consistent way |
| Are spelling and abbreviations kept as written? | A tidied transcript is no longer a record of the page |
| Do you use recognition software for a first draft? | If so, your questions about software apply to them too |
| How are our files returned and deleted? | You will want that in writing |
Whoever types, the transcript should follow fixed rules for gaps, strike-outs and insertions, and should be checked against the image by someone other than the typist. The post on how to transcribe handwriting to text sets out those rules and a checking routine you can hand to a service as your specification.
Software and people also combine well. A common split is to run everything through the tool that won your test so the whole file is searchable, and then commission human transcripts of only the pages the matter turns on. The test set tells you where that line falls for each writer.