Read the text out of a photo or scan

Usually one of three things. Handwriting is not supported — recognition is trained on printed type. A photo taken at an angle or in uneven light is much harder than a straight, evenly lit one. And a PDF that already has selectable text does not need this tool at all: copy it directly, or use PDF to JPG if you want the pages as images.

Both, depending on the page, and the result tells you which. Most PDFs already contain their text — anything exported from a word processor does — and where that is true the text is lifted out exactly, with no recognition involved and no chance of a misread. Only pages that are genuinely images, like a scan or a photo, are recognised, and those are a best guess.

For recognised pages: on a clean scan of printed text, very; on a phone photo of a curled receipt, much less. Recognition is a best guess rather than a transcription, so check names, numbers and anything that matters — a misread digit in an invoice total looks exactly as confident as a correct one. Pages read straight from the document need no such check.

Recognition uses the language to choose between characters that look nearly identical in print, and to weigh which words are plausible. Reading Spanish text with English selected produces noticeably worse results than picking Spanish.

Because it did not need reading. If a document already carries its text, there is nothing to render and nothing to recognise, so a long report finishes in about a second rather than the minutes a scan of the same length would take. Nothing was skipped or approximated — the opposite, the text is exact.

Because text is what people actually want when they cannot select anything in a scan, and it is honest about what recognition produces. A searchable PDF hides the guesses behind the original image, which makes mistakes harder to spot rather than easier.

Two hundred megabytes, and the first two hundred pages of a PDF. Each page is recognised separately, so a long document is genuinely a long job — but several documents can be read at once, so a queue moves quickly.