How AI reads handwriting in a PDF

Printed letters come from a font: every "a" on the page has the same shape. Handwriting has no font. The same letter changes from writer to writer and from one word to the next, letters join and overlap, lines slope and wander, and a quick "n" can look like a "u".

So the methods that read print well, which match each shape against the letters of known fonts, fail on handwriting. It needs models trained on handwriting itself.

Handwritten text recognition, or HTR, uses neural networks trained on large collections of handwritten lines paired with what they say. Rather than cutting a word into letters, the network reads a whole line and uses the context around each stroke to decide between shapes that look alike.

Newer models join an image model, which looks at the strokes, with a language model, which knows which words are likely. That is why they read cursive far better than older systems, and also why they can guess a plausible word when the writing is unclear.

How well it works depends on the handwriting, the language, and how much handwriting like it the model was trained on. Neat print-style writing can come out almost perfectly; a hurried signature or a doctor's note rarely does.

This converter is built for printed and typed text. Its OCR engine, Tesseract, is trained on print and is not meant for handwriting, so handwritten words on a scan usually come out wrong or are left out, while the printed parts of the same page convert as usual.

That makes it a good fit for a typed document with a few handwritten notes, or a printed form filled in by hand: convert it here and type the handwritten parts yourself. For a page that is all handwriting, use a handwriting service and check the result.