History of OCR: optical character recognition since 1913
Before a machine could read, someone had to teach it what a letter is. That took more than a hundred years, and every kind of reader brought its own way to be wrong. Which means: the history tells you what to ask of an OCR today.
Thirteen short scenes with a voice. Press play, or just scroll and read: the text is the whole story. Pictures and clips are AI-generated illustrations, not archive footage.
1913: the optophone turns letters into sound
The story begins in 1913, with an Irish physicist named Edmund Fournier d'Albe. Printed words were everywhere, but not for everyone. His optophone turned each letter into a chord of tones, so the blind could hear a page. It worked, but it needed a whole minute for a single word.
Sources
- Optophone (Wikipedia): invented by the physicist Edmund Fournier d'Albe. An early device read about one word a minute (shown in 1918 with Mary Jameson), later ones up to 60.
1951: Shepard's reading machine matches letters to stencils
Nearly forty years later, an American codebreaker named David Shepard built a reading machine in his spare time. It held every letter against stencils, one after another, until one fit.
Sources
- David H. Shepard (Wikipedia): a cryptanalyst at the Armed Forces Security Agency, he built the machine in his spare time; patent application of 1 March 1951.
- Shepard's Gismo (History of Information): the machine he called Gismo.
1954: Reader's Digest puts a reading machine to work
Three years later, Reader's Digest put one of his machines to work, turning typed sales reports into punch cards. But it knew one typeface only. If a letter looked a little different, the machine was lost.
Sources
- Reader's Digest and the reading machine (Incode): the first company to put one of the machines to work, turning sales reports into punch cards.
1965: the post office reads ZIP codes, 42,000 addresses an hour
By 1965, a machine in Detroit's post office was reading the ZIP codes on our letters. Forty-two thousand addresses an hour.
Sources
- Multiline optical character reader (Wikipedia): Detroit, November 1965. The machine reads only the line with city, state and ZIP code; 42,000 addresses an hour.
- Postal Reporter: distribution technology: the postal service on its distribution technology.
1968: OCR-A, a typeface made for machines
Still, machines stumbled over our letters. So in 1968, people changed the letters instead: OCR-A, a typeface no machine could mistake. Even our handwriting had to obey: neat block numbers, written for a machine.
Sources
- OCR-A (Wikipedia): made by American Type Founders; the character set is the standard ANSI X3.17 (1966).
1976: the Kurzweil reading machine reads books aloud
Eight years later, an American inventor named Ray Kurzweil turned it around. His machine read almost any printed typeface, and spoke it aloud. For the first time, a blind reader could hear an ordinary book in spoken words.
Sources
- Reading machine (Wikipedia): a machine that read ordinary printed books aloud.
2005: Tesseract, open source OCR for everyone
For decades, OCR was a tool for specialists: banks, post offices, big archives. Then Hewlett-Packard set its reading engine free: Tesseract. From then on, anyone could turn a scan into text, even at home. Some love it, some hate it. It runs on any computer, so it is still everywhere. But give it a crooked scan or a coffee stain, and the text falls apart.
Sources
- Tesseract (Wikipedia): developed at Hewlett-Packard from 1985 to 1994, released as open source in 2005 (its author Ray Smith says 2006).
2018: Tesseract 4 reads whole lines with a neural network
Twelve years after a German professor and his student published it at a small lab in Switzerland, Tesseract adopted their idea: a network that learns to read without being told where each letter is. No more cutting lines into letters: the machine read the whole line.
Sources
- Graves, Fernández, Gomez, Schmidhuber (IDSIA): Connectionist Temporal Classification, ICML 2006. A network learns whole lines without a mark for every letter.
- Tesseract 4.0.0 (GitHub): version 4.0.0 of 29 October 2018 brought the LSTM network and line recognition.
2023: Attention Is All You Need, and the large language models
This time it took only six years. A paper from Google, with a bold title: Attention Is All You Need. In 2023, its children arrive: the large language models. You hear them before you see them. Halls of machines, fans roaring like a storm. All that power, for a little chat window.
Sources
- Attention Is All You Need (arXiv): Vaswani and others at Google, NeurIPS 2017. The basis of the transformer, and of the large language models.
2025: LLM OCR reads pages as tokens, and some skip OCR
2025. Machines now read tokens, small pieces of words. The bold ones skip OCR altogether, and hand the raw picture of a page straight to the machine. A passport. A parcel. A file. Machines write. Machines read. Nobody sits at the desk. And mostly, it works. Some think OCR is finally solved.
Sources
- Paper on pages as image tokens (arXiv, 2025): a 2025 paper that squeezes pages into image tokens.
- heise test of the compressed mode: in a test on a magazine page, the most compressed mode wrote text that had nothing to do with the page.
- Study: Visual Merit or Linguistic Crutch? (arXiv): under strong compression the model guesses from language instead of reading; classic pipelines held up better (January 2026).
Is OCR solved? Hallucination, from a 2013 scanner to LLM OCR
Not quite. In 2013, a German computer scientist found numbers on his scan that were never on his paper. His scanner had compressed the image to save space. Compress a page too hard, and machines fill the gaps themselves. Every line reads perfectly. But not every line was on the page. Attention is all you need. To chat. Not to decide.
Sources
- CNBC on the changed numbers (2013): David Kriesel found that scanners swapped digits (6 and 8) through JBIG2 compression. In a floor plan with three rooms (14.13, 21.11 and 17.42 square metres) the scan showed 14.13 for all three. The maker confirmed it.
- PC World on scanners that alter numbers (2013): the same case.
OCR today: from text to context, in the archive
For a hundred years, machines gave us text. Now they give us context. But the most valuable paper never leaves the building: contracts, patient files, a company's memory. So the machine comes to it. Humans set aims. Agents work. In 1913, a machine needed a minute for one word. Today, one question sets a hundred years of paper to work.
OCR on your own server
The paper that matters most stays in the building. So we built an OCR that comes to it: it runs on your own server, hands back every page with a box and a confidence for every word, and marks what it cannot read instead of inventing it.
Production use, high volume or your own server?
Private use and tests are free. For a company in production, a server of your own or a pipeline built on your pages, send the kind of page and the volume, and get a written offer, or a clear no when the fit is not there.
