Amazon Textract alternatives: a self-hosted alternative to AWS Textract, or a hosted API

Tesseract is the usual answer when OCR has to run on hardware you control. It is Apache 2.0 software from the project repository, and its documentation says the install has two parts, the engine and the trained data for each language. It has no built-in window. It writes plain text, a searchable PDF, hOCR, TSV, ALTO or PAGE, and it does not return form fields or table cells, so that work stays with you.

Tesseract 5.5.4, installed locally with the matching language pack, got 55.3% of the Urdu page's characters wrong against 9.1% for FastOCR, and 20.4% of the Arabic page against 2.6%; lower is better. Tesseract's image-quality guide lists steps such as 300 dpi input and dark text on a light background, so a tuned setup may score better than this run. The Tesseract comparison has the full run.

Both are cloud APIs that answer with structured data rather than a finished file, so the differences sit in setup and scope. Google's setup guide asks for a Google Cloud project with billing and the Vision API enabled, and its PDF page says PDF and TIFF detection works on files in Cloud Storage and writes JSON to a bucket, for up to 2000 pages. AWS runs single pages through synchronous operations and larger documents through asynchronous ones, as its workflow page describes.

On languages, Google's language page says the document text endpoint auto-detects its full set and labels each language Supported or Experimental, with Urdu Experimental. AWS names more than 40 languages and then narrows them by operation. Textract's overview lists forms and tables as outputs, which a text-only comparison would miss. Neither service was run on our scans, so this page ranks neither. The Google Cloud Vision comparison covers its pricing and free tier.

Textract exists for structured extraction. Its overview lists text, forms and tables, and its analysis operations return key-value pairs and table cells as Block objects your code can read. FastOCR returns text in reading order and a searchable PDF; it does not return form fields or table cells as data. If your application needs those, stay on Textract.

Price is the other reason to stay. AWS publishes a separate per-page rate for each operation on its pricing page, checked 4 November 2026, so the operation you call sets the bill. At high volume on AWS, that rate table decides more than any feature list.

AWS's quotas page says Textract supports more than 40 languages, among them Arabic, Hindi, Chinese, Japanese, Korean and Hebrew. The same page limits Forms calls to English, Spanish, Italian, Portuguese, French and German, says other APIs are limited to English, and limits handwriting to English. Read the page for the exact operation you plan to call before you commit to a script.

Urdu, Persian, Bengali, Tamil and Thai are not among the languages that page names. FastOCR detects the language of each page itself, including pages that mix scripts, so there is no per-request setting. On our test scans it got 9.1% of Urdu characters wrong, 7.4% of Bengali and 2.2% of Tamil; lower is better. Those are single public-domain pages per script, not a guarantee for your documents.

FastOCR's website and API are hosted, so files are uploaded to FastOCR for processing; they are not an on-premise option. The free Mac and Windows apps read files offline on one computer and save a searchable PDF, text, Word and Markdown. They are public betas that recognise text and write files; they do not return forms or tables, and the accuracy figures on this page do not apply to them.

For a server deployment inside your own network, the open-source engine above is the route. You take on installing it, preparing the images and maintaining the servers, and you build any form or table extraction yourself.

The OCR API takes an image or a PDF and returns text, or a searchable PDF for scanned pages, so there is no Block list to walk. Create one API key and follow the quickstart. The Python SDK is fastocr-sdk on PyPI, the MCP server fastocr-mcp lets AI agents read documents through the same API, and LangChain and LlamaIndex loaders are published.

Every API account includes 20 free pages each month, and paid options are on the API pricing page. The trade-off is where the file goes: each document you send is uploaded to FastOCR for processing, which matters for regulated work.