How to read a scanned PDF when you are blind

The quick answer

A screen reader cannot read a scanned PDF because there is no text in it. The page is a photograph. Your screen reader is working correctly and finding nothing, which is why it often says nothing at all, or reads only the filename.

Four things can get you the content, roughly in order of how much structure they preserve:

1. Ask the sender for the original. Fastest and best when it works. The Word or source file has real text and real headings.
2. Your screen reader's built-in OCR. NVDA has an OCR command; JAWS has Convenient OCR. Free, immediate, and enough to find out what a page says.
3. A dedicated OCR or document tool. Better accuracy on poor scans, and some will attempt reading order.
4. An AI assistant that reads the page. The newest option, and the only one that reliably keeps tables and equations intelligible rather than flattening them.

If you only try one thing, try your screen reader's own OCR first. It is already installed.

Why a screen reader cannot read it

A PDF can hold two very different things. A digital PDF contains text characters, and often tags describing which text is a heading, which is a table cell, and what order to read it in. A screen reader handles that well.

A scanned PDF contains an image of a page. Someone put paper through a scanner, or photographed it. There are no characters, no headings, no reading order. There is a picture. A screen reader reads text, so there is nothing for it to do.

This is the difference between a document being untagged and being image-only. An untagged PDF has text but no structure, so a screen reader reads it in a jumbled order. An image-only PDF has neither. Both are common, and they need different fixes.

How to tell which you have: try to select text with the keyboard, or search the document for a word you can see is on the page. If nothing selects and nothing is found, it is image-only.

What OCR does, and what it loses

Optical character recognition turns pictures of letters into letters. It is decades old, it is good, and it is the foundation of every option here. What it does not do is recover meaning.

What OCR reliably gives you:

• The words, in most cases, on a reasonable scan.
• Rough line order on a single-column page.

What OCR usually loses:

• Tables. A table becomes a run of numbers. Nothing records which column a figure belonged to, so a row of results reads as a meaningless sequence.
• Multi-column layout. Two columns are often read straight across, interleaving two unrelated sentences.
• Mathematics. An equation becomes an approximation of its symbols. A fraction, an exponent, and a subscript all flatten into the same line of characters.
• Charts and images. There is nothing to recognise, so they are skipped in silence — and you are not told that something was skipped.
• Headings. Larger type is not marked as a heading, so you cannot navigate by structure.

That last group is why a scanned PDF can be "read" and still be unusable. The words arrive; the document does not.

The options, honestly

Ask for the original file

Under-used, and usually the best outcome. A short reply — "could you send the Word version, the PDF is a scan and my screen reader cannot read it" — often works, and it educates the sender. In an institutional setting it is also the request that changes future behaviour.

NVDA and JAWS built-in OCR

Both major screen readers can OCR the current document or screen. It is free, already installed, and takes seconds. Use it to find out what a page says. Expect it to struggle with the table and the equation, for the reasons above.

Dedicated OCR and document tools

Better recognition on faded, skewed, or handwritten-annotated scans, and some will make a real attempt at reading order. Worth having when scans are a regular part of your work rather than an occasional annoyance.

AI assistants

The distinction that matters is between recognising characters and understanding a page. A model that can look at the page can be asked what the table says, and answer by row and column. It can speak an equation as an equation. It can describe the chart that OCR silently skipped. It can be asked a question about the page rather than made to read it start to finish.

The honest limits: it can be wrong, and it will not tell you when it is. On anything consequential — a dosage, a deadline, a figure in a contract — verify independently. That is true of every tool on this list, but it is easier to forget with one that sounds confident.

Where Darvy fits

Darvy is in the last category. It is a voice-activated AI assistant that runs on your own Windows machine, alongside JAWS, NVDA, or Narrator rather than instead of them. You hold a key and ask.

On a scanned PDF it reads the page, keeps tables in their rows and columns, speaks mathematics properly, and describes charts and diagrams instead of skipping them. You can ask it a question about the page rather than listening to all of it.

It is free for 7 days and needs only an email address, so a scanned document you are stuck on right now is a reasonable way to test whether it helps. It runs on Windows 10 and 11. Results vary when a document or application exposes incomplete accessibility information, and AI-generated descriptions can be wrong — verify anything that matters.

See what it does or how to evaluate it against your own documents.

Frequently asked

Related reading on darvy.ai

• What Darvy can do
• Running Darvy alongside NVDA or JAWS
• What to put in an AT device demonstration program
• Darvy and JAWS, compared honestly