What is OCR?
Optical Character Recognition turns a picture of text into text you can search and copy. The useful question is not what it stands for — it is when it works, and it works far better and far worse than people expect, depending entirely on what is on the page.
The one number that matters
We measured two OCR engines on the same harness while building the handwriting tool: one metric (word accuracy, meaning one minus the word error rate), one set of samples, directly comparable. Tesseract is the classical engine behind most free OCR, including ours. TrOCR is a vision transformer — a fundamentally different approach.
| What is on the page | Classical OCR | Vision model |
|---|---|---|
| Printed text | 94.1% | 92.3% |
| Careful block printing | 49.4% | 83.1% |
| Semi-cursive handwriting | 1.1% | 71.8% |
| Real handwriting, unseen | 22.5% | 60.0% |
Read the third row again. 1.1% is not "poor accuracy" — it is noise. The engine that reads a printed page almost perfectly cannot read cursive at all. That is the single most useful fact about OCR and it appears on almost no page that explains it.
It also explains something that looks like inconsistency: a good OCR tool and a good handwriting tool are not the same tool, and should not pretend to be. They run different models with different costs — the vision model in that table is a 252 MB download against about 5 MB for the classical engine.
Why a scan comes out as gibberish
When OCR fails on printed text, it is nearly always the image rather than the engine. In rough order of how often it is the culprit:
- Resolution. Below roughly 300 dpi the letterforms stop being distinguishable — rn becomes m, cl becomes d. Rescanning at 300 fixes more failures than any other single change.
- Skew. Layout analysis runs before recognition and assumes lines are horizontal. A page scanned a few degrees off can produce a perfect character-by-character read of a completely scrambled reading order. If your output has the right words in the wrong sequence, this is why.
- A photograph instead of a scan. A phone photo of a page carries curl, uneven lighting and a shadow from your own hand. All three defeat the thresholding step that separates ink from paper.
- The wrong language. Engines are trained per language and will confidently map unfamiliar accented characters onto ones they know.
Recognised text is not the same as a searchable PDF
These get conflated constantly, and the difference decides which tool you want.
A text dump gives you the words and throws away the page: no layout, no images, no signatures. Fine for pasting a paragraph into an email.
A searchable PDF keeps the scan exactly as it looks and places the recognised text invisibly behind it, positioned over the matching words. It looks identical to the original; Ctrl-F works. For anything you need to keep — a contract, a deed, a medical record — this is almost always what you actually wanted, and it is the reason the OCR tool here hands back a PDF rather than a wall of text.
OCR used to mean uploading your documents
That was a real constraint, not a business model: the engines were too heavy for a browser, so the page went to a server. The consequence was that scanning a passport, a payslip or a medical letter meant handing it to a company you had not thought about.
That constraint is gone. Tesseract compiles to WebAssembly and runs in a tab in a second or two, and vision models run on WebGPU. Every OCR tool on this site works that way — the file is read by your own browser and never uploaded. It is also why they work with the network off.
Which tool do you want
- A scanned PDF you need to search. OCR a PDF — keeps the pages as they are and adds the text layer behind them.
- A photo or screenshot with text in it. Image to text, or screenshot to text.
- Handwriting. The handwriting reader — a different model, a real download, and honest about being a draft you correct.
- A page that scanned crooked. Straighten it first. It will do more for your results than switching tools.
Questions
- What does OCR stand for?
- Optical Character Recognition. It is the process of turning an image of text — a scan, a photograph, a PDF page that is really a picture — into characters a computer can search, copy and edit.
- Is OCR accurate?
- On clean printed text, yes: YourDevice measured 94.1% word accuracy for Tesseract on a printed control set, and the errors there are mostly punctuation and unusual names. On handwriting it collapses. The same engine scored 49.4% on careful block printing and 1.1% on semi-cursive — not "poor", but noise. Accuracy is not a property of OCR; it is a property of OCR applied to a particular kind of page.
- Can OCR read handwriting?
- Classical OCR essentially cannot read handwriting. Tesseract, the engine behind most free OCR, scored 1.1% word accuracy on semi-cursive handwriting in YourDevice’s tests. A different class of model — a vision transformer such as TrOCR, which is what /handwriting-ocr/ runs — reaches 71.8% on the same samples, which is a draft you correct rather than a result you trust. Anyone claiming reliable cursive recognition is selling something.
- Why does my scan come out as gibberish?
- Almost always resolution or skew. Below roughly 300 dpi the letterforms do not survive; a page photographed at an angle confuses the layout analysis before recognition even starts. Deskewing and rescanning fixes more failures than changing engine does.
- What is a searchable PDF?
- A PDF that still shows the original scanned image, with the recognised text placed invisibly behind it in the right positions. You see the scan; your search finds the words. That is different from a text dump, which gives you the words but loses the page.
- Does OCR need to upload my documents?
- Not any more. OCR ran on servers for years because it was heavy, but the engines now compile to WebAssembly and run in a browser tab. Every OCR tool on YourDevice works that way — image to text, handwriting to text, and searchable-PDF OCR — so your files are never uploaded, which for scanned contracts, medical letters and passports is the whole point.
Accuracy figures are our own, measured on a single harness with word accuracy (1 − word error rate) over the IAM handwriting database and a printed control set, August 2026. Tesseract was re-run through that same harness so every number in the table is directly comparable.