Coming soon: translate your documents between Indian languages and English.
Sanskrit OCR — Extract Sanskrit Text from Scans and Images
Upload a scanned Sanskrit PDF or an image of a Sanskrit page in Devanagari and get editable Unicode text back. Free, ad-free, no sign-up.
संस्कृतम् · Accepts PDF, JPG, PNG, TIFF up to 25 MB, a PDF up to 10 pages · Files deleted within 4 hours
Optical Character Recognition (OCR). Online & Free
Convert Scanned Documents and Images into Text
Drop PDF or image here. PDF, JPG, PNG, TIFF supported.
Up to 25 MB. A PDF can have up to 10 pages; split a longer one first and send the parts one at a time.
How to recognize text from image?
1
Upload your file
Click "Choose Files" or drag and drop your scanned PDF or image onto the upload area.
2
Select language
Choose the language of the text in your document from the dropdown for best OCR accuracy.
3
Download text
Click "Recognize" and download the extracted text file once processing is complete.
Unicode output
NFC-normalised text that pastes into any editor and stays searchable.
No ads, no sign-up
Free to use. Uploads removed within 4 hours.
Measured, not claimed
96.5% character accuracy on our published test corpus.
Legacy Sanskrit fonts and PageMaker
Older Sanskrit documents were often typed in a legacy font, which stores the keys pressed rather than Sanskrit letters, so copied text comes out as Latin letters. IndicPDF reads Sanskrit typed in Kruti Dev (010 and 714), DevLys and Chanakya. A PDF set in one of them is read from its font, here or in the Legacy Font tool, rather than by OCR.
For PageMaker, Word to PageMaker writes a Unicode Sanskrit Word file back in Kruti Dev 010, ready to place in PageMaker 7.
Sanskrit, Hindi and Marathi share the Devanagari script, and Auto detect never picks Sanskrit: a Devanagari page is read as Hindi or Marathi. The Sanskrit model is the better reader of Sanskrit. On our Sanskrit test page it got 3.54% of the characters wrong, the Marathi model 3.60% and the Hindi model 4.70%. So pick Sanskrit from the list when you know the page is Sanskrit.
Measured accuracy
On our ground-truth corpus the pipeline reads 96.5% of the characters of its one Sanskrit test page correctly (3.54% character error rate), 1,554 characters of prose. One clean digital page is a small sample, so read it as a direction rather than a rate. Old printed books, with their dense conjuncts and worn type, will do worse.
Where it still goes wrong
The errors on that page are the conjuncts Sanskrit uses more than Hindi does: ्न read as ्र, so शक्नोमि came back as शक्रोमि and स्वप्नम् as स्वप्रम्; ड्ड read as द्ध in उड्डीयमानः; ञ्च losing its first half in किञ्चित्; and ृ dropped from हृदये. Closing quotation marks were lost too. Handwriting and palm-leaf material is not supported; this reads printed Devanagari.
Frequently asked questions
स्कैन किए गए संस्कृत PDF से पाठ कैसे निकालें?
ऊपर PDF अपलोड करें और भाषा में संस्कृत चुनें। पृष्ठ यूनिकोड देवनागरी में पहचाने जाते हैं, जिसे आप कॉपी और संपादित कर सकते हैं।
Why should I pick Sanskrit when Auto reads Devanagari anyway?
Auto reads a Devanagari page with the Hindi or Marathi model. On our Sanskrit test page the Sanskrit model got 3.54% of the characters wrong against 4.70% for Hindi.
Is the visarga kept as a visarga?
Yes. All 52 visargas on our Sanskrit test page came back as ः, the Unicode visarga, and none as a colon.
What file types can I use for Sanskrit OCR?
PDF, JPG, PNG and TIFF, up to 25 MB per file. Scanned PDFs are rasterised page by page before recognition, so a PDF with no text layer works the same as a photograph.
How many pages can I send at once?
The OCR tool reads a PDF of up to 10 pages; a longer one is refused at upload, so split it with Split PDF and send the parts. PDF to Word reads up to 300 pages. A scanned page takes about 23 seconds to read, so a PDF with more than about 23 scanned pages is refused until it is split into parts.
What counts as a page?
One page of your file as it is stored, which is one side of a sheet. A book scanned two pages to a sheet counts one page for each sheet. A JPG or PNG is one page, and a TIFF counts every page it holds. All of them count towards the OCR tool's 10 pages and PDF to Word's 300, blank pages included. Only pages read by OCR count towards PDF to Word's limit of about 23 scanned pages a job: an empty page is not read, but a blank page in a scan is a picture of paper and is read like any other.
Is the output real Unicode text?
Yes. Output is standard Unicode, normalised to NFC, so it copies into Word, Google Docs or any editor and stays searchable. It is not an image of text and not a legacy font encoding.
Do I need an account?
No. The tool is free and requires no sign-up, no email and no payment. There are no advertisements on any page.
What happens to my files?
Uploads and results are stored encrypted and removed by a cleanup that runs every 15 minutes, so nothing is kept longer than 4 hours after upload. Files are processed to produce your result and are not used for anything else.
Should I use Auto Detect or pick the language myself?
Pick the language when you know it. Auto detection reads the script from the page and then verifies its guess, but a page with only a few lines, heavy noise, or mixed English gives it less to work with. Manual selection removes that uncertainty entirely.