Skip to content

LearnIngest & extract

Handwriting recognition

HTR · Handwritten text recognition

Transcribing handwritten text from an image. It is harder than printed OCR because letters join, vary between writers and follow no fixed font.

In DocSlurpGlossary entry

On this page

Handwriting recognition, also called handwritten text recognition (HTR), transcribes handwriting from an image. It is OCR without the easy parts: letters join into each other, every writer forms them differently, and the same writer’s “a” changes from word to word.

Models therefore lean heavily on context. Systems like TrOCR pair an image encoder with a language-model decoder, so a squiggle between “head” and “oom” is read as “r” because “headroom” is a word. Modern vision-language models do the same at page scale, and are often the most accurate option for handwritten notes and filled-in forms.

A concrete case: the inspector’s handwritten note on site-inspection.jpg reads “Gen tested 03/14, ran 17 min”. The words come back right. The digits are the risk: “17” versus “11”, “03/14” versus “03/19”. Context helps a model guess words; it gives no help with a number, and the model may confidently write the most likely-looking one.

The gotcha: language context is a double-edged tool. It corrects misread words and also “corrects” things that were written correctly but unusually: rare names, codes, misspellings that matter in a legal record. For forms, ask the model to transcribe digits exactly as written and not to infer missing ones, and route low-confidence fields to a person. See page routing.

In DocSlurp

In DocSlurp

There is no dedicated handwriting model. The optional page router labels images as handwriting and routes them to OCR with a higher quality floor (core/worker/src/routing/planner.ts), and the form-fields vision prompt tells the model to inspect every handwritten character and not infer digits (core/worker/src/tools/vision-runtime.ts).

Sources

Jump to