OCR that runs on your own computer

Turn scans and photos into searchable text.

Glyphwise reads scanned documents, PDFs and phone photos and gives you searchable PDFs, text and data. Every word comes back with its place on the page and a confidence score, so you know exactly which ones to check.

Download free trial Try it in your browser

Windows, Linux, macOS No account Works offline

A phone photo of an IRS Form W-4 with a blue box around every recognized word

✓823 wordsboxed on one photo

!31 flaggedfor a quick check

97.7%of words found on 38 test pages it never saw
0 bytesuploaded: everything runs on your machine
7languages, each with its own language model
3platforms: Windows, Linux and macOS

How it works

From a crooked phone photo to text you can search.

Real output: an IRS form photographed at an angle in uneven light, read by the model that ships in the download.

A phone photo of a W-4 form, tilted and shaded

Drop in your files

Scans, PDFs, TIFFs or photos. Tilted, sideways and shaded pages are straightened and evened out first.

The same photo with a box around every word: blue when sure, orange when not

Every word is read

Each word gets a box and a confidence score. Blue means sure; orange means worth a second look.

A paragraph of the form selected as text in a PDF viewer

Search, copy, export

Get a searchable PDF you can select and copy from, or text, JSON, hOCR and ALTO for your systems.

Features

Everything you need to digitize paper.

Know which words to check

Every word has a confidence score. Send only the doubtful ones to review, as a spreadsheet, instead of proofreading whole pages.

Close-up of form text with a blue box around every word

Nothing leaves your computer

No cloud, no telemetry, no activation server. Contracts, medical records and tax forms stay on your machine, even air-gapped.

Searchable PDFs

Your scan, unchanged, with an invisible text layer: search, select and copy like any digital document.

Text selected in a searchable PDF made from a photo

Forms, reports and registers

Columns, tables and form boxes are found and read in order. Tested on IRS forms, NIST reports and the Federal Register.

NIST report title page, every word boxed

Federal Register page in three columns, every word boxed

Seven languages

Each with its own language model that corrects likely misreadings.

EnglishFrançaisDeutschEspañolItalianoPortuguêsNederlands

Fast on ordinary PCs

Uses every processor core, and an NVIDIA card if you have one. 38 test pages in 13 seconds on a laptop with an NVIDIA card.

Multi-core CPUCUDAOpenCL

Built into your software

A library with ready wrappers, a local HTTP service and a command line.

Python.NETJavaNode.jsGoC

Formats

Feed it the files you already have.

PDFs with a text layer are read directly, so born-digital documents come back exact. Scanned pages and phone photos are repaired first: crooked, sideways and upside-down pages are set straight and shadows and speckles are cleaned away.

Input

PDFText layers read directly; scanned pages recognized. Archive.org-style MRC scans supported.
TIFFMulti-page; uncompressed, LZW, Deflate, PackBits, CCITT G3/G4, JPEG.
JPEG 2000, JBIG2Including the formats used inside library and archive PDFs.
PNG, JPEG, GIF, BMP, PNMPhone photos of pages (straightened and evened out) and screenshots.

Output

Searchable PDFYour scan with an invisible text layer: search, select and copy.
JSONPages, blocks, lines and words, each with a box and a confidence score.
hOCR, ALTOThe standard formats for digitization, library and archive systems.
Plain text, review listText for indexing, and a spreadsheet of the words worth checking.

Measured, not claimed

Fewer errors on real documents.

We test on public-domain government documents, forms, phone photos and handwritten archives the models never saw in training, and run other engines on exactly the same files.

Test set	Measure	Glyphwise	Tesseract 5.4	PaddleOCR 3.7	TrOCR large
Government documents2,000 lines from 8 held-out documents	character errors	0.33%	1.47%	—	—¹
Full pages, forms included38 pages: Federal Register, IRS forms, NIST report	words found (F1)	97.7%	94.1%	94.4%	—
Same pages, words with a letter or digitleaves out punctuation, dot leaders and form rules	words found (F1)	98.3%	97.7%	98.0%	—
Second page set53 pages from 27 other held-out documents	words found (F1)	97.4%	94.0%	—	—
Phone photos of pagesthe 38 pages photographed at an angle, in uneven light, slightly blurred	words found (F1)	96.6%	90.9%	—	—
Handwritten manuscriptsBentham papers, 18th–19th century, 424 lines	character errors	14.0%	74.6%	—	12.7%
Modern handwriting, phone photosGNHK notes and forms, 300 lines	character errors	44.9%	71.1%	—	32.2%
Speedthe 38 pages above, laptop with an NVIDIA GPU	total time	13 s	21 s	—²	—

Lower is better for errors, higher for words found. Pages rendered at 300 dpi; every engine with its standard English model, run as a separate program on the same files and scored the same way (— = not run on that set). Glyphwise numbers are for the model that ships with the product.

Where Glyphwise leads: printed documents and forms, and phone photos of them. Words found counts every word the page's text layer holds; much of the full-page lead comes from punctuation, dot leaders and form labels the other engines often leave out, so the row with letters and digits only is the closer contest. Where it does not: handwriting. TrOCR, a 560-million-parameter transformer that needs a graphics card, reads modern handwriting markedly better and old manuscripts slightly better; Glyphwise also skipped a few lines it could not read (6 of 424 and 11 of 300), which flatters it a little. The photo copies are generated from the test pages, so their ground truth is exact.

¹ TrOCR's printed-text model writes capital letters only, so it is not scored on documents. ² PaddleOCR ran on the processor only, where it took minutes per page.

Languages

Seven languages, each with its own model.

English, French, German, Spanish, Italian, Portuguese and Dutch. Pick one with --lang, per request in the HTTP service, or in the demo.

Language	Read with the English model	With its own model	Fewer errors	Lines without an error
Englishthe default	—	0.22%	—	92%
French`--lang fr` · try it	1.53%	0.91%	−41%	64%
German`--lang de` · try it	1.36%	0.24%	−82%	87%
Spanish`--lang es` · try it	1.52%	0.44%	−71%	80%
Italian`--lang it` · try it	1.02%	0.64%	−37%	64%
Portuguese`--lang pt` · try it	2.16%	0.54%	−75%	68%
Dutch`--lang nl` · try it	0.86%	0.15%	−83%	92%

Character errors on 500 lines per language from real documents: five EU regulations and directives (data protection, digital services, copyright, services, artificial intelligence), 100 lines from each, published as PDFs in every EU language and rendered at 300 dpi; the PDF's own text is the ground truth. None of these documents was used in training. Source: EUR-Lex, © European Union. Other Latin-alphabet languages are read with the English model.

Three ways to run it

From one folder of scans to your own application.

Command line

Batch-convert folders: one searchable PDF and text file per scan, processed in parallel. Add --gpu to use an NVIDIA card.

glyphwise models\line.bin scans ^
  --out-dir out --formats pdf,txt

Local service

An HTTP service for your document system or scanner workflow. It stays on your network.

curl --data-binary @scan.pdf ^
  "localhost:8080/ocr?format=alto"

Library

Ready wrappers for Python, .NET, Java, Node.js and Go, and a small C API for anything else. See the code.

doc = ocr.read("scan.pdf")
print(doc.text)

For developers

OCR inside your own software.

Every word comes back as a typed object with its text, confidence and position; errors arrive as your language's exceptions. One loaded model serves many threads.

from glyphwise import Glyphwise

with Glyphwise("models/line.bin") as ocr:
    doc = ocr.read("scan.pdf", lang="fr")
    for w in doc.words():
        if w.confidence < 0.8:
            print("check:", w.text, w.bbox)

using Glyphwise;

using var ocr = new GlyphwiseOcr("models/line.bin");
Document doc = ocr.Read("scan.pdf", lang: "de");
foreach (var w in doc.Words())
    if (w.Confidence < 0.8)
        Console.WriteLine($"check: {w.Text} {w.BBox}");

import site.glyphwise.*;

try (var ocr = new Glyphwise(Path.of("models/line.bin"))) {
    Document doc = ocr.read(Path.of("scan.pdf"));
    for (var w : doc.words())
        if (w.confidence() < 0.8)
            System.out.println("check: " + w.text());
}

const { Glyphwise } = require("glyphwise");

const ocr = new Glyphwise("models/line.bin");
const doc = ocr.read("scan.pdf", "es");
for (const w of doc.words())
  if (w.confidence < 0.8) console.log("check:", w.text, w.bbox);
ocr.close();

ocr, err := glyphwise.Open("models/line.bin")
if err != nil { log.Fatal(err) }
defer ocr.Close()
doc, _ := ocr.ReadFile("scan.pdf", "it")
for _, w := range doc.Words() {
    if w.Confidence < 0.8 { fmt.Println("check:", w.Text) }
}

Also: plain JSON, hOCR and ALTO output, the C API underneath (include/glyphwise.h), and examples for each language in the download. Building it into an application is part of the Developer licence.

Your documents stay yours

Nothing leaves your machine.

Contracts, medical records and tax forms are exactly the documents that should not be uploaded to someone else’s cloud. Glyphwise never connects to the internet.

No cloud processingRecognition runs entirely on your computer or server.

No telemetryThe software sends nothing: no usage data, no crash reports.

No activation serverYour licence is a signed file checked offline. It works on air-gapped machines.

Built in-houseThe recognition engine and models are our own, with no third-party OCR inside.

A quiet website, tooNo cookies and no advertising trackers; visits are counted anonymously with GoatCounter. See the privacy notice.

Clear termsAnnual licences, no automatic renewal, 30-day refunds. See the terms of sale.

Pricing

Simple yearly licences.

Each licence runs for one year and includes every update released in that year. No automatic renewal. Prices in US dollars, before applicable tax.

Desktop

For one person converting scans

$149per workstation, per year

Command-line program
Searchable PDF, JSON, hOCR, ALTO, text
GPU acceleration
Local HTTP service
Library for your own software

Server

For a team or a document workflow

$990per server, per year

Everything in Desktop
Local HTTP service for your internal systems
Unlimited users inside your organisation
Email support
Library for your own software

Developer

For building OCR into internal applications

$2,900per site, per year

Everything in Server
Library with wrappers for Python, .NET, Java, Node.js and Go
Up to 5 developers at one site
Priority email support

30-day refunds · Licence file by email · Works offline

Shipping Glyphwise inside your own product, or offering it as a hosted service? That needs a distribution licence. Write to us with what you are building and roughly how many customers it will reach. Contact sales

Free trial

Try it on your own documents first.

Download the evaluation package for Windows, Linux or macOS and unpack it. No installer, no account.
Run glyphwise on a few of your own scans. Without a licence it reads up to 3 pages per run, with every feature switched on.
When you buy, your licence file arrives by email. Put glyphwise.lic next to the glyphwise program and the limit is gone.

Download the free trial

Windows 10/11 x64, Linux x64 (Ubuntu 22.04 or newer) and macOS 12 or newer (Apple silicon and Intel). See what's new. Each package includes the model, the user guide, the licence agreement and third-party notices.

Questions

Before you buy

Which languages does it read?

English, French, German, Spanish, Italian, Portuguese and Dutch, each with its own language model that corrects likely misreadings (measured here). Other languages written in the Latin alphabet are read with the English model, which works but makes more mistakes on accented words. Non-Latin scripts (Greek, Cyrillic, Arabic, Chinese, ...) are not supported.

Does it need a graphics card?

No. It runs on any modern x64 processor and uses all its cores. An NVIDIA card (or any OpenCL card) makes large batches faster; the results are identical either way.

How good is it on handwriting?

Much better than engines built only for print, but handwriting is still hard: on 18th-century manuscripts about one character in seven is wrong, and casual modern handwriting photographed on a phone is harder still. Use the confidence scores and the review list to find the words to check.

What happens when my licence expires?

The software goes back to evaluation mode (3 pages per run) until you renew. Files you already converted are yours and are not affected.

Can I get a refund?

Yes. If Glyphwise does not work for you, ask for a refund within 30 days of purchase. See the terms of sale.

Do you support Linux or macOS?

Yes. There are packages for Windows (x64), Linux (x64) and macOS (one build for Apple silicon and Intel), with the same features and the same licence file. GPU acceleration uses CUDA or OpenCL on Windows and Linux; on a Mac recognition runs on the CPU.

Can it read photos taken with a phone?

Yes. Pages photographed at an angle are found and flattened, sideways and upside-down photos are turned upright, and shadows are evened out. For best results photograph the whole page, flat, in even light, against a darker background.

Where is the documentation?

The user guide covers installation, the licence file, the command line, the HTTP service and the C API. It is also in the download.

Ready to read your paperwork?

Download the free trial, or try it in your browser right now.

Download free trial Open the demo