OCR vs. handwritten text recognition
| OCR | Handwritten text recognition (HTR) | |
|---|---|---|
| Built for | Printed and typed characters | Connected, variable handwriting |
| Works well on | Typed minutes, printed newsletters, forms | Consistent, legible hands; scripts it was trained on |
| Struggles with | Any handwriting; faint carbons | Unusual scripts, abbreviations, faded ink, crowded layouts |
If a tool calls itself "handwriting OCR," it's usually HTR under the hood. The label matters less than how it does on your pages.
Why old handwriting is hard
- Scripts change. Nineteenth-century copperplate, clerks' hands, and twentieth-century cursive look very different, and letterforms that are obvious to a human reader confuse models.
- Abbreviations and conventions. Registers and ledgers are full of shorthand: "do." for ditto, abbreviated names and months, superscripts.
- Ink and paper. Faded iron-gall ink, bleed-through, foxing, and water damage remove the contrast recognition depends on.
- Layout. Ruled ledger columns, marginal notes, and entries written across each other break the reading order.
- Many hands. A register kept for sixty years may have dozens of writers.
Your options
Phone apps and general handwriting-to-text tools
Fine for a handful of modern, legible pages. Rarely enough for historical volumes: results vary, and most give no indication of which words they guessed.
Dedicated handwritten text recognition platforms
Platforms such as Transkribus are built for historical documents and can be trained on a specific hand. They're powerful and suit projects with someone willing to learn the tool and correct output.
Human transcription
Volunteers or paid transcribers remain the most accurate route for difficult material, and many historical societies run transcription projects. It's slow, and it needs consistent rules for abbreviations and uncertain readings.
A service with recognition plus review
The middle path: machine recognition on everything, human review on the pages that need it, and a record of which is which. That's how Boxchive works.
Getting the best results from photos
- Light evenly. Diffuse daylight or two lamps at angles; no flash glare.
- Hold the phone parallel to the page so lines aren't skewed.
- Fill the frame with one page, and include the edges so nothing is cut off.
- Focus on the writing, and retake anything blurry — recognition can't recover detail the photo didn't capture.
- Keep page order consistent so entries that run across pages stay together.
Don't let search silently skip handwritten pages
The real risk isn't a typo in a transcription. It's a search that returns nothing because a register was never read — which looks exactly like "we don't have that." Whatever tool you use, keep track of which pages were transcribed confidently, which were partial, and which weren't read at all.
Boxchive flags handwritten, faded and damaged pages for human review, and your archive records what was confidently transcribed and what needs a second pair of eyes. More on how this fits into search: what makes an archive searchable.