Your documents stay yours
The recognition model sits on your hard disk. Files are never uploaded, never stored by us, never used to train anything. For anyone handling client papers that is not a feature — it is the whole point.
We would rather ship a tool that genuinely works than five that almost do. Every one runs entirely on your own computer — no account, no upload, no document ever sent anywhere.
Scanned documents into editable Word — entirely offline.
Twenty PDF tools that run on your machine, not a website.
Opens files that will not open — Word, Excel, PowerPoint, PDF.
Protect code so it can't be copied — Python, JavaScript, PHP, .NET, Java, Go.
Lock any file, folder or mailbox — offline. Only you can open it.
Both numbers on the right are measured on an ordinary laptop with no graphics card — the machine already on your desk.
And then you proofread it, because you will have made mistakes.
Words it was unsure of are already underlined for you to check.
So it goes to a free website, and the file leaves your computer.
And the redaction actually removes the text, rather than covering it.
Do either twice a month and you have bought back a working week a year. Each licence costs less than one of those afternoons.
One invoice, its exact text known in advance, damaged nine ways: rotated 3° and 10°, blurred, noisy, low contrast, shrunk, skewed in perspective, shadowed, and saved as a heavy JPEG. Each damaged page is read, and the words that come back are matched against the known text.
94% is the mean across all nine. The worst case is 91% — the perspective-skewed page. Timing is 7.1 seconds a page on an ordinary laptop CPU with no graphics card, which is what most people will run it on.
The 98.7% is one real deed, scanned at 300 DPI: 3.27 MB in, 45 KB out, still 300 DPI and still searchable. Measured by the compression code itself, not estimated. A colour page compresses less, which is why the app inspects each page and tells you the result before it runs.
Different engines suit different damage. ARIA OCR ships more than one and picks per page, so a document that defeats one is still read by another. Your scans will differ from ours; the method will not.
Every figure on this site comes out of the code or a test run, not a marketing meeting. Where there is a trade-off, we say what it costs.
We tested four recognition engines on the same page and shipped the fastest one that lost no text: 0.73 seconds against 17.4 for the common default, on a third of the memory. That is why it runs on a four-year-old laptop instead of needing a graphics card.
| What we ship | 0.73 s |
|---|---|
| Second-fastest we tested | 7.2 s |
| The common default | 17.4 s |
| Slowest we tested | ~20 s |
Same page, same machine, model already loaded. How ARIA OCR works →
smaller — ARIA PDF took a real 300 DPI scan from 1.78 MB to 0.018 MB at full resolution, still searchable.
The honest limit: that mode discards colour, so it is used only on text pages. See the measurements →
The usual way to convert or edit a document is to hand it to a server you have never heard of. For a holiday photo that is fine. For a client file, a medical record or anything under NDA, it is the whole problem.
The recognition model sits on your hard disk. Files are never uploaded, never stored by us, never used to train anything. For anyone handling client papers that is not a feature — it is the whole point.
Four gigabytes of memory is comfortable. No graphics card. No server to rent, no queue to wait in. Most software assumes new hardware; this one does not.
Running handwriting is largely beyond it, and stylised headings sometimes garble. Both are written on the product page, before you pay. Seven free days to check we are telling the truth.
Install it, point it at the worst scan you own, and see what comes back. No card, no commitment.