Image to Text OCR: 120+ Languages, Hebrew, Arabic, CJK

Turn images and scanned PDFs into text. Give it links or upload files, pick the languages (or let it detect the script), and get back clean text in reading order, every line with its position on the page and a confidence score.

It reads 123 languages, including right-to-left Hebrew, Arabic, Persian and Urdu, Chinese (simplified and traditional), Japanese, Korean, Hindi, Thai, Russian and Greek. It runs the open-source Tesseract 5 OCR engine inside the actor: no third-party API, no login, no proxy. Your files are not sent anywhere else.

Price: $0.0015 per image or page with up to 1,000 characters of text. No custom start fee. Pages with no text found are free.

What you can use it for

• Screenshots and images to text: chat screenshots, slides, posters, product labels, memes, social media images.
• Scanned documents: scanned PDFs, faxes and multi-page TIFF scans, letters, forms, old books and archives.
• Right-to-left and Asian scripts: Hebrew and Arabic come out in logical reading order (right word first), ready for search or translation.
• Layout work: line and word boxes in pixels for highlighting, redaction, or cropping a region of the page.
• AI and RAG pipelines: feed the text of scanned pages to an LLM, a search index or a translation step.

How to use it

1. Paste image or PDF links into Image or PDF URLs, or upload files under Upload images or PDFs. (The sample link is a public test image from the Tesseract project. Remove it if you only upload files.)
2. Pick the Languages of the text. Choose Auto to detect the script of each page and turn rotated pages upright.
3. Optional: for PDFs and multi-page TIFFs, set Pages (for example 1-3, 7) and Max pages per file. Set Boxes to lines, lines and words, or none.
4. Run it, then download the results as JSON, CSV or Excel, or read them through the API.

Input

Field	Required	Example	Notes
urls	one of urls or files	["https://example.com/scan.jpg"]	Up to 1,000 links per run
files	one of urls or files	uploaded in the Console	Stored in your own Apify key-value store
languages	no	["heb"], ["eng", "fra"], ["auto"]	Tesseract codes. Default English
pages	no	"1-3, 7, 10-"	PDFs and multi-page TIFFs. Empty = all pages
maxPagesPerFile	no	100	1 to 2,000
boxes	no	"lines"	lines, words (lines plus words) or none

{
    "urls": [
        "https://example.com/receipts/receipt-0193.jpg",
        "https://example.com/archive/scanned-letter.pdf"
    ],
    "languages": ["eng", "spa"],
    "pages": "1-5",
    "boxes": "lines"
}

File types: JPEG, PNG, WebP, TIFF (all pages), GIF and BMP (first frame), and PDF (each page is rendered at 300 DPI and read). HEIC photos from iPhones need converting to JPEG first. Files up to 50 MB.

Output example

One dataset item per image, or per page of a PDF or TIFF. This is the real result for the sample image (lines cut to two):

{
    "url": "https://raw.githubusercontent.com/tesseract-ocr/test/232ff181c66516116ec0e84c4963f70de15050fd/testing/phototest.tif",
    "fileName": "phototest.tif",
    "fileType": "image",
    "fileSizeBytes": 38668,
    "page": null,
    "pageCount": 1,
    "width": 640,
    "height": 480,
    "rotated": 0,
    "detectedScript": "Latin",
    "languages": ["eng"],
    "text": "This is a lot of 12 point text to test the\nocr code and see if it works on all types\nof file format.\n\nThe quick brown dog jumped over the\nlazy fox. The quick brown dog jumped\nover the lazy fox. The quick brown dog\njumped over the lazy fox. The quick\nbrown dog jumped over the lazy fox.",
    "lineCount": 8,
    "wordCount": 60,
    "charCount": 285,
    "confidence": 96.2,
    "lines": [
        {
            "text": "This is a lot of 12 point text to test the",
            "confidence": 96.3,
            "box": { "left": 36, "top": 92, "width": 544, "height": 30 }
        },
        {
            "text": "ocr code and see if it works on all types",
            "confidence": 95.1,
            "box": { "left": 36, "top": 126, "width": 582, "height": 31 }
        }
    ],
    "error": null,
    "readAt": "2026-10-31T17:16:46.167854+00:00"
}

• text: lines end with a newline, paragraphs are separated by a blank line.
• confidence: the engine's own score from 0 to 100, per page and per line (per word with Boxes = words). Below about 60 the text is usually worth a look.
• box: pixels from the top left corner of the upright page. For PDFs, the page at 300 DPI. Very large images are read at up to 25 megapixels, and their boxes are mapped back to the original size.
• A link that cannot be read (404, not an image, password-protected PDF) gets a row with an error message and is not charged. A page where no text is found gets a row with empty text and is not charged either. The other files in the run carry on.

Pricing

Event	Price	When
Image or page read	$0.0015	Each image or page, including up to 1,000 characters of text
Extra 1,000 characters	$0.0007	Each further 1,000 characters (or part) on the same page
Actor start	$0.00005	Apify's standard start event, once per run per GB of memory (once at the default 512 MB)

Most screenshots, photos and receipts have under 1,000 characters, so they cost $0.0015 each: $1.50 per 1,000 images. A typical letter page with 2,500 characters costs $0.0015 + 2 x $0.0007 = $0.0029. A dense page of small print with 7,000 characters costs $0.0057. Characters are counted in the text field you get back (charCount), spaces and line breaks included. Pages where no text is found and files that fail are free.

The extra event exists because reading time grows with the amount of text: a page of dense small print takes about ten times longer to read than a screenshot.

Before each page the actor checks your run's maximum charge and stops when the next page would go over it. A run never goes over the limit you set. It also stops cleanly shortly before the run's timeout, so a long batch ends with the pages it read rather than a timed-out run.

Languages

Pick one or more in the Languages list. Each extra language makes reading slower, so pick only those that appear.

• Auto: detects the script of each page (Latin, Hebrew, Arabic, Cyrillic, Greek, Han, Japanese, Korean, Devanagari, Thai and about 20 more) and uses its main language, and turns pages that are upside down or on their side upright. For Latin-script text Auto reads English, so for French, German, Spanish, Portuguese or Italian pick the language yourself to get the accents right. Auto needs a few lines of text to tell the script: on a single word or one short line it reads English and detectedScript is null, so pick the language for short text.
• Right to left: Hebrew (heb), Arabic (ara), Persian (fas), Urdu (urd), Yiddish (yid), Pashto (pus), Sindhi (snd), Uyghur (uig), Syriac (syr), Dhivehi (div).
• Asian: Chinese simplified (chi_sim) and traditional (chi_tra), Japanese (jpn), Korean (kor), each also in a vertical-text variant, plus Thai, Vietnamese, Khmer, Lao, Burmese, Hindi, Bengali, Tamil, Telugu and other Indic scripts.
• European: all major ones, plus historical models such as German Fraktur (frk), Middle English, Middle French, Old Spanish and Old Italian.

Tips and limits

• Best on clean, flat text: scans, screenshots and photos of documents taken straight on. Text on curved objects, street signs at an angle, handwriting and stylized logos read poorly; this engine is built for printed text. A lone word in a large empty image can come back empty (and is not charged).
• Resolution matters: text should be at least about 20 pixels tall. Tiny text in a small screenshot reads better if you upscale the image first.
• Rotated pages: Auto turns pages by 90, 180 or 270 degrees when it is confident. Slightly skewed scans are read as they are.
• Speed: a screenshot takes a few seconds; a dense scanned page takes about a minute on the default 512 MB. For big batches, give the run more memory (up to 4 GB) at the same price, since you pay per page: in our test, 10 dense scanned pages took 526 seconds at 512 MB and 134 seconds at 2 GB.
• PDFs with a text layer are read by OCR too. If your PDF already has selectable text, a PDF-to-text tool is faster and exact.
• Links must be public http or https URLs. The actor respects each site's robots.txt: if a site disallows automated access to a path, that file gets an error row and no charge.

FAQ

Which OCR engine is this? Tesseract 5 with its fast LSTM models, a widely used open-source OCR engine (Apache 2.0). It runs inside the actor, so there is no per-call fee to an outside API and your images stay in the run.

How accurate is it? On clean printed text it is close to exact: the sample page above comes back word for word, and our dense scanned test pages averaged a confidence of 95. The confidence score tells you which pages and lines to check. Photos with glare, blur, perspective or handwriting do much worse. Try a few of your own files first: a small run costs a fraction of a cent.

Does it handle Hebrew and Arabic properly? Yes. Words come out in logical order (the first word you read is first in the text), without the invisible direction marks some OCR tools leave in. Hebrew with vowel points (nikud) reads with lower confidence than plain text.

Can I call it from code, a schedule or an AI agent? Yes. Use the Apify API, a schedule, an integration such as Make or Zapier, or an MCP client. It runs with limited permissions and needs no access to your other Apify data.

Do you keep my files? The actor reads each file during the run and writes the results to your own run storage. Uploaded files stay in your own key-value store, under your control.

Something looks wrong or a file failed. What do I do? Open an issue on the Issues tab with your run link and, if you can share it, the file.