A scan stores pictures of words rather than selectable characters. Image to text can recognise English text from those pages on your device.
OCR estimates characters from shapes. Small type, skew, blur, low contrast and handwriting can cause errors even when a page looks readable to a person. Use a clear scan and pay particular attention to reference numbers and decimal points.
Illustrative example: recognise a scanned invoice, then compare the supplier name, invoice number and total against the page before pasting them into a spreadsheet. A recognised 0 and O or 1 and I can look plausible but change the meaning.
This workflow produces plain text, not a searchable PDF or an editable recreation of the page. It does not preserve the visual layout of tables. The OCR assets need to load before an offline run, and long scans may exceed browser memory.
The pages are drawn at about 150 dots per inch and then read, so a clear, straight scan gives much better text than a crooked or shadowed one. The reading works for printed English, and it does not read handwriting.
The tool shows an average confidence. Above about 88 percent the text is dependable, between 70 and 87 it is reasonable but needs proofreading, and below 70 you should find a better scan. Check names, dates and numbers against the original, because those are where a wrong letter matters most.
The current implementation loads English recognition data only.
No. The output is copied text or TXT, separate from the original PDF.
It depends on the number of pages. The first time you use it, the reading engine of about 4 MB is downloaded from this site and then kept on your device.