Cursive Translator

Tools and tradeoffs

Can AI read cursive handwriting? Test it on a page you know

Can AI read cursive handwriting from your page? This page does not answer that question on your behalf or describe another product. It states what this site's own request flow consists of, provides a blank comparison worksheet, and sets out a test protocol for a page whose reading you already know. Record the observations from your own test alongside its date and input.

A worksheet, not a ranking. Last reviewed .

What follows: the request this site makes and what comes back from it; an inspection list for your own image and page; a comparison worksheet you fill in yourself; a repeatable test protocol; privacy questions to ask before you upload anywhere; and a review procedure for machine output. If you want to try this site's tool first, the cursive translator takes one image and writes whatever text is returned into an editable box, with no account.

What this page states, and what it refuses to state

Start with what is verifiable here, because it is the only thing this page can vouch for. This site is not a recognition system. It is an upload form and a relay.

  • The request. Your browser posts one base64 image to /api/cursive-to-text. That endpoint sends one request to the Google Cloud Vision images:annotate API with feature DOCUMENT_TEXT_DETECTION and the single language hint en.
  • The response. One text field is taken from the reply, trimmed, and written into the editable box. Confidence values, bounding boxes and any other structure in the reply are not read and not shown.
  • No tuning. All three upload pages use the same request-building code, endpoint, feature type and language hint. There is no model choice, script setting, per-page OCR configuration or post-processing.
  • The request checks. Accepted types JPG, PNG, WEBP, GIF, BMP, TIFF; a 6 MB client-side check; an in-memory threshold of 10 requests per IP per minute; and KV counter thresholds of 100 per IP and 3000 site-wide per UTC day.

Now what this page will not do. It will not tell you how well that request performs on your page, because this site publishes no measurement of it. It will not describe how any other product behaves — what a chat assistant does with an image, how another OCR service is built, whether either repeats itself on a re-run — because those are other companies' systems, they change, and a tool site's account of a competitor is exactly the kind of claim you should refuse from us as readily as from anyone else.

What is left is more useful than either: a way to find out. The rest of this page is inspection lists, a blank worksheet and a protocol. You fill them in from your own document.

Behaviours to test for, rather than categories to trust

Instead of relying on category labels, run the observations below on the tool you are using — including this site's tool — and write down what you saw. The worksheet does not predict how long the checks will take or what they will show.

  • Send the identical file twice, an hour or a day apart. Diff the two outputs. Record whether they are character-for-character identical, differ in a few words, or differ in formatting and length.
  • Count what came back that is not transcription. Preambles, apologies, summaries, headings the page did not have, notes about what it could not read. Record how much you had to strip out before the text was usable.
  • Check whether anything is marked as uncertain. Look for per-word confidence, highlighting, brackets, or any signal at all. Write down what you found; on this site, the answer is nothing, because the page reads only the text field.
  • Compare spelling and punctuation against the page. Take three words your document spells unusually and check whether they came back as written or in a modern form. Do the same for capitalisation and for line breaks.
  • Look for structure. If your page has columns or a form, check whether the output preserves any of that or returns one running stream.
  • Time the whole task. From the file on your desktop to corrected text where you need it, including sign-in, prompt writing, waiting, re-asking, copying out and cleaning up.
  • Note what the tool required of you. An account, a subscription, a prompt, an app install, a page visit. Record it; it is part of the comparison.

Each item above asks for an observation rather than asserting a property of another tool. Write the date, input and observed output next to it so the entry remains tied to the test you actually ran.

Inspect your image and your page before you compare anything

Record properties of the submitted file separately from the returned output. Go through both columns with the image open at full zoom and write down what you find. These are inspections, not diagnoses.

Inspect the image

  • Sharpness. Zoom to 100% and look at a single letter. Are the stroke edges crisp or soft? Write down which lines are softest.
  • Angle. Lay a straight edge along a written line on screen. Does the line stay under it, or curve and converge?
  • Light. Compare a word at the top left with one at the bottom right. Is the paper the same brightness in both? Note any shadow or bright patch, and where it falls.
  • Show-through. Hold the original to the light or check the reverse. List the marks in your image that come from the other side of the leaf.
  • Resolution. Note the pixel dimensions and the file size. Then note whether it is the original capture or a screenshot, export or re-share of one.
  • Damage. Mark the folds, stains, tape, trimmed edges and holes, and list which words they cross.

Inspect the page

  • Words you cannot read yourself. List them by line number, at full zoom. This records your unresolved words for the comparison; it does not establish that another reader would find the same words ambiguous.
  • Proper names. List every personal and place name. These are the words you will check one at a time, against another source where you have one.
  • Numbers. List every date, sum, age and quantity so each can be checked directly against the image rather than accepted because it looks plausible.
  • Abbreviations and unfamiliar forms. List the contractions, superscripts and letterforms you do not recognise, and note the line each appears on.
  • Non-paragraph layout. Note the columns, marginalia, interlinear insertions, struck-through text and anything written across an earlier page.
  • Cost of an error. Write down, in a sentence, what a wrong word on this page would cost you. That sentence sets how much review the page gets.

Keep both lists. They are what you will compare any output against — and the left-hand list is also the set of things you can change and re-submit as an experiment, one at a time, judging the result yourself.

One property of this site's own output is worth restating here, because it governs the whole review: what arrives in the box carries no marks. The page reads only the text field of the response, so nothing distinguishes a word that matches your page from one that does not. That is why the review procedure below exists.

For archival material, the companion worksheet on reading old cursive sets out provenance fields and a notation to record what you could not resolve.

A comparison worksheet you fill in yourself

The middle column states properties of this site that are fixed in its own code. The right-hand column is deliberately blank: fill it in for each tool you try, from your own observation, with the date you observed it. This page ranks nothing and recommends nothing.

Properties to record for any tool you try. Middle column: this site, from its code. Right column: yours to complete.
Property to check This site Tool you are testing — fill in
Account or sign-in required None. No account system exists in the code.  
Cost, and any cap No payment control on this site. The endpoint checks configured thresholds of 10 requests per IP per minute, 100 per IP and 3000 site-wide per UTC day.  
Accepted input One image per request: JPG, PNG, WEBP, GIF, BMP, TIFF, up to 6 MB. No PDF, no batch.  
Who processes the image A Cloudflare Pages Function on this site forwards it to Google Cloud Vision.  
What the output arrives in An editable <textarea>, with copy and download buttons acting on its current contents.  
Uncertainty marked in the output Nothing. Only the text field of the response is read; confidence values are not shown.  
Structure preserved This page displays one text field and does not render response coordinates.  
Same file submitted twice — identical output? Not stated here. Run the test and record what you get.  
Non-transcription text to strip out None added by this site; the returned string is trimmed and written as-is.  
Wall-clock time for the whole task Yours to measure.  
Words wrong, dropped or invented on your test page Yours to measure.  

The last four rows require observations from your own run and cannot be filled in from this site's code. Run the protocol below, write the answers in, and date the sheet so each entry identifies its test conditions.

A repeatable test protocol

A procedure for producing your own numbers. Keeping the input and recorded steps fixed reduces known differences between attempts, but it does not establish the cause of every observed difference.

  1. Choose a page you can grade

    Pick a page whose reading you already have, or transcribe it by hand first and keep that as your reference. Otherwise the worksheet compares outputs without an independently checked reference. A page of running hand can be used for one test; record its line count rather than treating a particular length as sufficient.

  2. Take one image and use that same file everywhere

    Record the capture conditions — page position, light, camera angle, framing and shadow — and submit the identical file to every tool. Re-shooting changes the input and introduces another possible source of differences.

  3. Ask for the same thing each time

    Where a tool accepts instructions, write one fixed wording and reuse it character for character: transcribe verbatim, preserve original spelling and line breaks, mark anything unreadable rather than supplying a word. Where a tool accepts no instructions — as this site's form does not — record that in the worksheet, because it is part of the comparison.

  4. Count three things separately

    Against your reference, count words that differ, words that are missing or added, and proper names and numbers that differ. Keep the third count separate: it answers a different question from the first two, and you may care about it far more.

  5. Time the whole task, not the request

    Start the clock at the file on your desktop and stop it when corrected text is where you need it. Include sign-in, prompt writing, waiting, re-asking, copying out and stripping anything you did not want.

  6. Repeat once, a day later

    Submit the same file again and diff the two outputs. Record whether they are identical; this is one direct observation relevant to the stability row.

Record the results somewhere plain. A sheet like this is enough:

INVENTED EXAMPLE — a blank sheet, not results, with made-up
sample details to show the shape.

Sample: letter, Hand A, 12 lines, 118 words. One photo, 2026-08-07.
Instruction (where accepted): "Transcribe verbatim. Keep original
spelling and line breaks. Mark unreadable words, do not guess."

Tool             Wrong  Dropped/  Names &   Total   Identical
                 words  invented  numbers   time    on re-run?
-----------------------------------------------------------------
Tool A             .        .        .        .         .
Tool B             .        .        .        .         .
Own reading        0        0        0        .        n/a
A blank sheet with invented sample details, not results. This site publishes no accuracy figures of its own. A number from a different sample is not evidence about the page you are testing here.

Privacy questions worth asking before you upload

A handwritten page can carry a full name, an address, a medical detail, a family matter, a signature. Before you send one anywhere — this site included — put these questions to whatever you are about to use, and answer them from that service's own current statement rather than from anyone's summary of it.

  • Is the image stored, and for how long? Processing an image and retaining it are separate things. Look for an explicit statement rather than inferring one from silence.
  • Who else processes it? Ask which other services the image is passed to, and read their terms too, because those then apply to your document.
  • Is it used to train anything? Ask about the specific plan you are on rather than the company's general position.
  • Can you delete it, and does deletion reach anything downstream? A delete control in one interface says nothing about copies held elsewhere.
  • Is an account required, and what does that account hold afterwards? An anonymous single upload and a stored history attached to your identity are different exposures.
  • Does this particular page justify any of it? For a will, a medical letter, or anything covering another living person, transcribing by hand yourself remains an entirely legitimate choice.

The answers for this site, so you can hold it to the same standard: an image you upload is posted to a Cloudflare Pages Function here and forwarded to Google Cloud Vision. The site code has no image or transcript database and does not intentionally persist either item. It writes numeric KV values under global and per-IP daily counter keys, daily operational-counter keys and alert-throttle keys; those entries contain neither the image nor returned text. No account exists. What Google or Cloudflare do with a request is governed by their own current policies, which this site does not summarise. The full statement is on the privacy page, alongside the terms.

This page will not characterise how any other provider handles your data. Their policies are theirs to state and yours to read; they change; and a tool site claiming to know what a competitor does with your files is exactly the kind of assertion you should refuse — from us as much as from anyone.

Reviewing machine output

Whatever produced the text, do this before you use it. The aim is to turn an undifferentiated block into something that shows where your own verification stopped.

Read against the image, not against sense

Put the output beside the image and check word against word, line by line, keeping the original's lineation so you can always find your place. Do not accept a passage because it reads smoothly: fluency is a property of the sentence, not evidence about the ink.

Verify every name, number and date individually

Check each one against the page on its own, and against an independent document where you have one. Mark each as verified or not; do not carry an unverified name forward into anything else.

Mark what you could not confirm

Replace anything you cannot verify with a bracketed note — [unclear], or [unclear: reading?] where you have a candidate — rather than leaving the machine's word standing unexamined. The reading cursive worksheet sets out the notation for gaps, insertions and abbreviations.

Where the output and your reading of the image disagree, record both

Write your own reading and keep the machine's as an alternative in brackets. You are holding the document; the output is a string. Do not overwrite what you can see with what was returned.

Say where the text came from

Note at the top that it began as machine output, which tool produced it, on what date, and who reviewed it. That header gives later readers an explicit record of the review you performed.

If a tool shows per-word confidence, you can use it to order your review, but do not treat a high value as verification: whatever it describes, it is not a comparison against the ink on your page. This site shows no confidence values at all — the page reads only the text field of the response.

Questions

Can AI read cursive handwriting from a photo?

This page does not answer that on your behalf. What it can tell you is what this site does with your image — one Google Cloud Vision request, one text field returned into an editable box — and how to find out what any tool does with your page: run the protocol above on a page whose reading you already have, and count the differences yourself.

Should I use this site or a general AI assistant?

This page states no ranking and no comparison of that kind. The worksheet above lists the properties worth recording; the middle column is filled in for this site from its own code, and the right-hand column is yours to complete from your own tests.

Why does machine output sometimes contain words that are not on the page?

This page does not explain another system's internals. What it states is the consequence for you: what arrives here carries no confidence values and no marks, so nothing in the box distinguishes a word that matches your page from one that does not. Compare every word against the image, and mark what you could not confirm.

What can I change and test?

The image is the part you control. Change one thing — the light, the angle, the crop, the flatness of the paper, the resolution you saved — submit again, and compare the two results against your own reading of the page. Repeat with one change at a time. That comparison is about your page; this site makes no prediction about the outcome.

Can it read very old or ornate handwriting?

This site states nothing about how any document reads. Every OCR submission uses one feature type and one language hint, with no historical, script or language setting; request metadata and counters can still differ. If your material is archival, the reading old cursive worksheet sets out how to record provenance, transcribe exactly and mark what you cannot resolve.

How accurate is it, in numbers?

This site publishes no figure. Any single percentage would belong to somebody else's sample, photograph and moment. Run the test protocol above on your own page and produce your own three counts.

What happens to my image on this site?

Your browser posts it to /api/cursive-to-text; that endpoint sends it to Google Cloud Vision and returns the text field of the response. The site code does not write your image or the returned text to any store of its own. No account exists. What Google does with it is governed by Google's own current policies, which this site does not describe. Details are on the privacy page.

Can I trust machine output for legal or genealogical purposes?

Not unverified. Site policy: verify every name, date and number against the original page, mark what you could not confirm, and record that the text began as machine output and who checked it.

What do I do with a stack of pages in the same handwriting?

Submit them one at a time — the form takes one image per request — and note the endpoint's configured checks of 10 requests per IP per minute and 100 per IP per UTC day. Review each page against its image before moving on, and keep a list of the readings you settled for your own later comparison. Start at the cursive translator, or handwriting to text if your pages are forms and notes.

Keep reading