How the recognition works
This guide covers how text recognition works here, which export to choose, how to turn a scanned page into a searchable PDF, and what to do when the text comes out wrong. Recognition runs on your device with Tesseract, an open-source OCR engine, so the image and the text stay with you.
Tesseract looks at the image, finds blocks, lines and words, and reads each word with a neural network trained for the language you selected. It returns the text, the position of every word on the image, and its own confidence for each word. The Accuracy figure shown on the result is the average of those confidence scores. It is the engine's own estimate, not a measured error rate, so a high number does not guarantee that every figure on an invoice is right.
Each language has its own trained data file. The first time you use a language, the browser downloads its file and caches it, so later runs start without that download. If you add a second language, both files are loaded and the engine considers both for every word.
Step by step
- Drop a photo, scan or screenshot on the page. JPG, PNG, WebP and HEIC all work. You can add several images; each is read separately.
- Click Edit on an image to crop it to the text and straighten it with rotate. Removing photos, logos and empty margins reduces stray characters.
- Choose the Text language. Same as site language uses the language this site is shown in. Pick the actual language of the document if it differs.
- If the page mixes two languages, for example Hindi with English product names, choose a Second language.
- Press Extract text. A progress bar shows the recognition; the first run for a language also includes its download.
- Read through the text box and correct mistakes in place. The word count updates as you type, and Undo my changes restores the original recognised text.
- Copy the text, or download it as TXT, Word or PDF.
Which export to choose
| Export | What you get | Includes your corrections | Best for |
|---|---|---|---|
| Copy text | Plain text on the clipboard | Yes | Pasting into an email, chat or form |
| TXT | A plain UTF-8 text file | Yes | Notes, code, search tools, archiving |
| Word (.docx) | One paragraph per line on an A4 page; right-to-left paragraphs for Arabic | Yes | Editing further in Word, Google Docs or LibreOffice |
| PDF (searchable) | The original image as the page, with the recognised words as an invisible, selectable layer | No, it keeps the text as recognised | Archiving scans you want to search and copy from later |
The Word file contains text only. Bold, font sizes, columns and tables from the original are not rebuilt; table rows come out as lines of text. The searchable PDF uses a standard Latin font for its hidden layer, so the PDF button only appears when every selected language uses the Latin alphabet.
Worked example: a scanned invoice to a searchable PDF
Suppose your scanner saved an English invoice as a JPG of 2480 × 3508 pixels, which is an A4 page at 300 dpi. You want a PDF you can later search for the invoice number.
Crop away the scanner's blank border with Edit, leave the language on English and press Extract text. Check the result against the scan, paying particular attention to numbers: OCR engines commonly confuse 0 with O, 1 with l, and 5 with S, and a misread digit in an amount is easy to miss.
Press PDF. The page in the PDF is the scan itself, stored as a JPEG, with each recognised word placed invisibly over the spot where it appears. In a PDF reader you can search for the invoice number, select a line and copy it, while the page still looks exactly like the paper.
One detail to know: the PDF sizes the page as if the image were scanned at 150 dpi. A 300 dpi scan therefore opens as a page twice as wide and tall as A4. On screen this makes no difference once the reader fits it to the window; when printing, choose fit to page. A scan made at 150 dpi comes out at A4 size exactly.
If you corrected a word in the text box, that correction goes into the TXT and Word files but not into the PDF, which always carries the text as recognised. For an invoice, keep the PDF as the searchable archive copy and use the corrected TXT or Word file when you need the figures elsewhere.
Supported languages
| Group | Languages | Searchable PDF |
|---|---|---|
| Western European | English, German, French, Spanish, Italian, Portuguese, Dutch, Irish, Maltese | Yes |
| Nordic and Baltic | Danish, Swedish, Finnish, Estonian, Latvian, Lithuanian | Yes |
| Central European | Polish, Czech, Slovak, Slovenian, Croatian, Hungarian, Romanian | Yes |
| Other Latin script | Indonesian | Yes |
| Cyrillic and Greek | Russian, Bulgarian, Greek | No; use TXT or Word |
| Arabic and Devanagari | Arabic, Hindi | No; use TXT or Word |
The hidden layer of the searchable PDF uses a font with a Western European character set. Some letters used in Central and Eastern European languages fall outside it and are left out of that layer, so a search for a word containing them may not find it. The TXT and Word exports keep every character.
Troubleshooting poor results
- No text was found: the language may be wrong, the text may be too small, or the image may be mostly picture. Crop to the text, check the language and try again.
- Faint or low-contrast scans: run the image through Filters & adjust first. Black and white with a threshold, or a contrast boost, often turns grey pencil-light text into clean black letters that recognise far better.
- Photos taken at an angle: shoot the page straight on and fill the frame. Curved lines near the spine of a book are read less reliably than flat text.
- Tiny text in a screenshot: text only a few pixels tall is hard to read for any engine. Taking the screenshot at a higher zoom level works best; upscaling 2× first can also help a little.
- Columns come out in the wrong order: multi-column layouts such as newsletters are sometimes read across the columns. Crop each column and run them one at a time.
- Handwriting: Tesseract is built for printed text. Neat block capitals may come out partly right; joined-up handwriting usually does not.
- Mixed languages read as nonsense: add the second language. Two languages are slower than one, so only add a second when the page really needs it.
- The download fails on first use: the language file could not be fetched. Check your connection; once it has downloaded, it is cached.
Memory and large batches
The recognition engine takes around 30 MB of memory while the tool is open, plus the language data, and it is released when you leave the page. Very large photos take longer and need more memory; on a phone, a 12-megapixel photo of a single receipt is better cropped to the receipt first. The free plan accepts up to 50 images per batch and files up to 200 MB; Pro allows 500 images and 1 GB per file. Each image produces its own result and its own exports, so a 10-page document becomes 10 text results rather than one combined file.
How this compares with cloud OCR
Cloud OCR services upload your image and run it on their servers. Their larger models are often better at handwriting, complex tables and photos taken in poor light, and some rebuild the page layout in Word. If your document is not sensitive and has a complicated layout, that can be worth the upload.
Invoices, bank letters, medical forms and ID documents are exactly the papers people prefer not to upload. Here the image and the recognised text stay in your browser, there is no account, and nothing is kept anywhere else. The costs are a one-time download per language and speed that depends on your device. With Pro, the installed app works offline once the languages you need have been downloaded.