Plan a pilot box
Menu

Text recognition for print and handwriting

Printed pages, typescripts and handwriting turned into text people can search, with the error rate measured on your own pages before anyone relies on it.

An image of a page is not searchable. Someone looking for a name in a box of letters still has to open every image and read it. Text recognition turns those images into text: optical character recognition (OCR) for printed and typed pages, and handwritten text recognition (HTR) for manuscripts. We run both, measure how well they did on your material, and correct the text to the level the collection needs.

What we recognize

  • Printed books, pamphlets and reports. Modern OCR handles clean type well. Old typefaces, the long s, ligatures and faded impressions need checking.
  • Typescripts and carbon copies. Typed letters and reports are usually reliable; carbons and mimeographs are not, because the letters bleed.
  • Newspapers and clippings. Columns, headlines and advertisements have to be split into reading order before the text makes sense.
  • Handwriting. Letters, diaries, ledgers and registers. Recognition quality depends on the hand, the ink and the paper, so we measure it on a sample before quoting the work.
  • Forms and tables. Ledgers and registers are read cell by cell so that a name stays attached to its date and amount.

How the work runs

  1. Sample. In the pilot box we pick pages that represent the collection: the clean ones, the faded ones and the worst hand in the box.
  2. Transcribe a reference. A person transcribes the sample exactly, which becomes the yardstick for every engine and model we try.
  3. Run and measure. We run recognition and compare the output with the reference. That gives a character error rate (CER) and a word error rate (WER) for each kind of page.
  4. Choose the correction level. With the numbers in hand, you decide where full correction is worth paying for and where uncorrected text is good enough to search.
  5. Process and check. The full batch runs, a correction pass follows where agreed, and a random sample of every batch is measured again before delivery.

What an error rate means for searching

The character error rate is the number of characters the machine substituted, missed or added, divided by the number of characters in the reference. A rate of 2 percent sounds small, but on a page of 1,200 characters it is about 24 mistakes, and a mistake inside a surname can hide that person from a search. Word error rate is higher than character error rate, because one wrong letter spoils a whole word.

Example (fictional). Two pages of a ledger hold 2,400 reference characters. Recognition substitutes 38, misses 12 and adds 9: 59 errors, a CER of 2.46 percent. Across a 1,200-page batch at six seconds per fix, full correction would take about 59 hours. The OCR accuracy calculator runs the same arithmetic on your own counts.

Because of this, we report error rates per batch and per kind of page, not one number for the whole collection. A batch of typed minutes and a batch of pencil diaries should not share a score.

Correction: full, targeted or none

Level What a person does When it fits
Full correction Reads every line against the image and fixes every error Text that will be quoted, published or used as evidence
Targeted correction Fixes names, places, dates and numbers, and lines the engine was unsure about Large collections where search is the main use
Uncorrected Delivers recognition as it came out, labeled as uncorrected Printed material with low error rates, or a first pass to decide what deserves more

Uncorrected text is still useful. It finds most mentions of most words. What matters is that everyone downstream knows it is uncorrected, so we label it in the files and in the catalog record. Our AI policy explains how machine output is marked until a person has checked it.

What you receive

  • Plain text for each page, in reading order.
  • ALTO XML for each page, the layout standard maintained by the Library of Congress, with the position of every word so a viewer can highlight search hits on the image.
  • Searchable PDFs where you want a single file per item.
  • An accuracy report for each batch: sample size, CER, WER and the correction level applied.

These files load into most collection systems and into the collection search we build. Names, places and dates found in the text feed the next stage, AI cataloging.

Engines and models

We use more than one engine, because no single one is best for every kind of page. For print, open-source engines such as Tesseract do well on clean type. For handwriting, platforms such as Transkribus train recognition models on transcribed pages from the same hand or period. We choose by measured results on your sample, not by brand, and we tell you which engine produced which text.

Your images and transcriptions are not used to train a third party’s general model without your written permission. When a custom model is trained for your collection, the agreement says who may use it afterward.

Start with your hardest pages

The quickest way to know what recognition will do for your collection is to try it on your own material. Send a pilot box with the pages you are most worried about, or read the fictional example collection to see what the text and records look like at the end. For background, the Reading Room’s text recognition section covers OCR and handwriting in depth.

Send one box first.

A pilot box shows you the images, the recognized text and the catalog records before you commit a whole collection. Tell us what is in the box and what you want people to find.

Measure recognition on a pilot box

Studio and contacts

24 South Weber Street, Suite 225
Colorado Springs, CO 80903
(719) 244-8450sales@dastok.com