PDF to text
Typed pages are copied out exactly. Scanned pages are read with OCR — and the page tells you which was which.
Up to 100 MB and 40 pages. Pages that already hold real text are copied straight out. A scanned page is read with OCR, which takes about seven seconds for a full page of print.
Checking whether the text reader is already on this device
0%The short answer
Every page is asked for its own text first. A PDF written by a computer already contains the words, so they come out exactly and instantly, with no OCR involved at all. A scanned page holds a photograph of paper instead, so that page — and only that page — is rendered at 180 dpi and read by Tesseract here in the tab, a 4.6 MB download the first time. The panel says which pages came out which way, because words a machine guessed deserve a second look and words copied out of the file do not.
Drop in a PDF and get its text back, whether it was typed or scanned. Free, no account, nothing to install — and it tells you page by page which words were copied out exactly and which ones a machine had to read.
The formula
page's own text → if there is none → render at 180 dpi → OCR that page → text
Worked examples
- A report exported from Word
- exact text, no OCR needed
- A contract with one signed page
- OCR on that page only
- A photographed page, held square
- very nearly every word
Reference table
| The PDF | How the text comes out | Speed |
|---|---|---|
| Made by Word, LaTeX or a browser | Copied out exactly | Instant |
| Scanned or photographed | Read by OCR, with a confidence figure | About 7 seconds a page |
| Scanned, then made searchable elsewhere | Copied out exactly — including that tool's mistakes | Instant |
| Part typed, part scanned | Whichever applies, page by page | Only the scans wait |
How to use it
- 01Drop the PDF inNothing is uploaded. pdf.js opens it inside the tab, one page at a time, and the text appears as it goes rather than all at the end.
- 02Watch which pages need readingPages with real text finish instantly. A scanned page is rendered at 180 dpi and read, which took about seven seconds a page on a dense A4 scan when we measured it, and the progress line says which page it is on.
- 03Check the summary under the textIt says which pages were copied out exactly and which were guessed at by OCR, with the confidence figure for those. Copied text needs no checking; guessed text does.
Also searched for
Searches that land here include pdf to text, pdf ocr, extract text from pdf, scanned pdf to text, copy text from pdf and pdf to txt.
Two ways into a PDF
The word 'PDF' covers two unrelated files
One is a document: characters, fonts, positions, everything a program needs to draw the page and everything you need to copy it. The other is a photograph of paper wrapped in a PDF container, which is what a scanner, a phone app or a fax gateway produces. They open in the same viewer and look identical on screen, and they could not be more different to get text out of.
The first takes milliseconds and is exact. The second needs OCR, takes about seven seconds for a dense page, and produces a best guess. Most tools treat every page as the second kind, which is why they are slow on files that did not need to be, and why they hand back guessed text where exact text was sitting right there.
How each page is decided
Every page is asked for its text first. If what comes back is under thirty characters once whitespace is removed, the page is treated as scanned — that threshold is well below any real page of prose and comfortably above the stray 'Scanned by…' stamp a scanner app leaves behind. Only then is the page rendered, at 2.5 times its natural size, which works out at 180 dpi and is the point where OCR accuracy stops improving enough to be worth the memory.
Mixed documents are the normal case
A signed contract is typed except for the signature page. A report has scanned appendices. An old file has been re-saved so many times that some pages carry text and others do not. Deciding page by page means those files come out with the typed pages exact and only the scanned ones guessed at — and the summary line under the text tells you which pages fell into which group, so you know precisely how much of the result to double-check.
Questions people ask
Why were some pages instant and others slow?
Because they are two completely different jobs. A PDF made by Word, LaTeX, a browser or almost any program contains the actual characters, positioned on the page — pulling them out is a lookup, and it is exact. A scanned PDF contains a photograph of paper and no characters at all, so the only way in is to render that page and read it like a picture. This tool checks each page and does the cheap thing wherever it can.
My PDF was already 'searchable', so why is the text wrong?
Because someone ran OCR on it before you did, and this tool is copying out what they got. When a scanner or Acrobat makes a scan searchable, it adds an invisible text layer over the image — mistakes included, permanently. That layer is real text as far as any PDF reader is concerned, so it is copied out verbatim rather than re-read. If the result looks mangled, the mangling happened years ago. Exporting the pages as images and running them through the image reader will give you a fresh attempt.
Do columns and tables survive?
Roughly, and only roughly. Line breaks and the spacing between columns are preserved, which is usually enough to see the shape of a table and to fix it up in a spreadsheet. A two-column academic paper will interleave in places, and a complicated form will lose its alignment. Nothing here reconstructs a table into rows and cells — if that is what you need, the text is a starting point rather than an answer.
Why does it stop at forty pages?
Because reading a scanned page happens on your own processor, and it is not instant. We timed a full A4 page of dense print at about seven seconds, so forty of them is a wait of several minutes — the text does appear page by page as it goes, rather than all at the end, but it is still a wait. A sparse page such as a receipt or an invoice takes well under a second, so the real limit depends on your document. If your file is longer than forty pages the page says exactly how many it left out, and pages that already hold text are effectively free — a 400-page typed report is not what this ceiling is protecting you from.
Is my document uploaded?
No. There is no server here that takes files. Two things are downloaded to your device — pdf.js, which opens the file, and the OCR engine, which only arrives if a page actually needs it — and the document itself never leaves the tab. For anything with a name, an address or an amount on it, that is the difference that matters.
What about a password-protected PDF?
A PDF that needs a password to open cannot be read here, and it fails with a message rather than pretending. A PDF that merely forbids copying is a different thing: that restriction is a flag the reader is asked to respect, not encryption, and pdf.js ignores it — so those files do work. Whether you should is between you and whoever set the flag.