Hindi OCR — Image & Scanned PDF to Text, Numbers and English Kept

Hindi documents are rarely only Hindi. A government circular carries dates, amounts, file and phone numbers, and English words such as office names, "Annexure" or "Loan Account", often inside a table with ruled borders. The Hindi model of Tesseract, the open-source OCR engine, reads Devanagari well but has no Latin letters, so English comes out as made-up Devanagari. It also loses the digit 1, so 12 becomes 2 and 1800 becomes 800, and a table with borders often comes back as noise.

This tool reads every page twice, once with the Hindi model and once with the English model, and keeps for each word the reading that has the evidence: Devanagari from the Hindi one, numbers and English words from the English one. Before reading, it erases the ruling lines of tables; only long, thin lines are erased. The text keeps the line order of the page. You can drop a photo, a screenshot or a scanned PDF.

We measured it on 14 pages of real PDFs published by Indian government accounts offices, which mix Hindi with English, numbers or both, against the words in the PDFs' own text. The Hindi model alone got 19% of the numbers and none of the English words right. Erasing the table lines alone lifts the Hindi model to 56% of the numbers, still with no English. With both steps, this tool got 92% of the numbers and 91% of the English words. On the same pages turned into rough 150 dpi scans, tilted and noisy, it got 77% of the numbers and 86% of the English words. We tuned the last rule on those 14 pages, so we then ran 26 other pages from the same offices that played no part in the tuning. Those pages are in English, many with tables, so they test the numbers and English words only: 89% of each came out right, and 79% of the numbers and 84% of the words on the rough scans.

The Hindi itself is as good as the Hindi model gets, not perfect. On those circulars, 62% of the Hindi words matched the PDF text exactly, against 60% for the Hindi model alone, and on real scanned book pages about 9 characters in 100 came out wrong. Clean, printed pages do much better: on our own Hindi test pages, 99.8% of the Hindi letters were right, digits in running text went from 80% to 99.3% right, and on three test tables with borders the character error fell from 81% to under 1%.

The honest limits. Handwriting is not read reliably. The result is plain text: tables come back as rows of words separated by spaces, not as a grid, and pages with two columns may mix them. Faint or blurred scans lose accuracy, so scan at 200 to 300 dpi if you can. Reading twice takes time: in our tests from about 4 to 15 seconds per page on a computer, longer for dense tables and on a phone, and the first use downloads about 4 MB. Marathi, Nepali and other languages written in Devanagari use the same letters, but we have not measured them. Everything runs in your browser, so the document is never uploaded.