Hindi OCR — image to text

Hindi · measured on this model

What the Hindi model got wrong in our test.

Hindi is written under a continuous headline (shirorekha), with vowel signs above, below and before the letters they belong to. That is why most Hindi errors are small ones: a dot, a sign, one stroke of a conjunct. The checklist below starts with what we saw the Hindi model do on our own test pages, then lists look-alike letters worth a second look. It is not a complete list of every possible error.

How we tested: we wrote 12 lines of Hindi (384 characters, none copied from a third party), drew them in two Noto fonts (SIL OFL), degraded them in 7 ways and ran the same recognition engine and Hindi file this page uses, with its default settings. We aligned the output to the text letter by letter. This is a small synthetic test that shows real failure modes. It is not an accuracy figure for your documents, and real scans and photos will behave differently. A script in the source code reproduces it byte for byte; the exact versions are under Sources at the end of this page.

Character errors by condition

• Clean render, sans, 36 px

  0.0% character errors

  Engine confidence 94

• Clean render, serif, 36 px

  0.3% character errors

  Engine confidence 95

• Small text, sans, 18 px

  0.0% character errors

  Engine confidence 94

• Photo-like: sans 28 px, 1.5° tilt, blur, noise, JPEG q45

  0.0% character errors

  Engine confidence 94

• Tiny text, sans, 13 px

  0.8% character errors

  Engine confidence 90

• Poor phone photo: sans 22 px, 3° tilt, blur σ1.4, heavy noise, JPEG q30

  2.1% character errors

  Engine confidence 92

• Faded print: serif 20 px, grey ink on tinted paper, blur σ1.1

  0.0% character errors

  Engine confidence 93

Proofreading checklist

Hindi: check these before you trust the text.

Items marked “observed” come from the test above, with the number of the 7 conditions in which each one appeared. Items marked “look-alikes” are a general proofreading habit for this script that we did not measure.

1. फ़ क़ ज़

   Observed · 1 of 7 conditions

   01The nukta dot

   Under the poorest photo conditions in our test, a letter with a nukta was read as the plain letter plus a half-ra: फ़ॉर्म came back as फ्रॉर्म and क़ानून as क्रानून. The other nukta words in the same test (ख़बर, ज़िंदगी, फ़ैसला, बड़ी, लड़की, पढ़ाई) were read correctly. Search the result for ्र where you expect a dotted letter, and check every ड़ and ढ़.

   • फ़ॉर्म → was read as फ्रॉर्म
   • क़ानून → was read as क्रानून

2. आँ आं

   Observed · 1 of 7 conditions

   02Chandrabindu and anusvara

   The moon-and-dot sign was read as a different letter in one test: आँसू became ऑसू (the model produced the vowel ऑ). Check every word where a nasal sign sits above a long vowel, and decide whether you meant ँ, ं or ॉ.

   • आँसू → was read as ऑसू

3. र्थ ष्ट्र

   Observed · 1 of 7 conditions

   03Parts of long words and conjuncts

   At very small type, a vowel sign disappeared inside a long conjunct word (विद्यार्थियों became विद्यर्थियों); in a blurred photo a conjunct lost a component (राष्ट्र became राष्र). Spell-check does not always catch these because the result can still look like a plausible word.

   • विद्यार्थियों → was read as विद्यर्थियों
   • राष्ट्र → was read as राष्र

4. १२ २

   Observed · 1 of 7 conditions

   04Devanagari digits

   The digits in a date were misread in two different conditions: १२ became ९२ in tiny text and १? in a serif render. Proofread every number, date, amount and reference code against the original.

   • १२ → was read as ९२
   • १२ → was read as १?

5. शर्मा ने

   Observed · 2 of 7 conditions

   05Lost word spaces

   Two words were joined into one (शर्मा ने became शर्माने) in the mildly degraded photo and in faded print. Because the headline joins neighbouring words, a short gap is easy for the model to lose; look for words that look too long.

   • शर्मा → was read as शर्माने

6. ब व घ ध भ म

   Look-alikes · not measured

   06Look-alike letters

   These pairs differ by a small loop or stroke. We did not measure them separately, so treat this as a general habit: re-read names, places and numbers that contain them, and compare ि with ी (the short and long i) at the start of a word.

7. । |

   Look-alikes · not measured

   07The danda and the vertical bar

   The sentence-ending danda looks like a vertical bar or a letter l. Not measured on Hindi pages by us; in the Marathi test the danda was read as ) and ! in two conditions, so check sentence ends.

Photographing Hindi

Tips for photographing Hindi documents.

• Keep the headline level

  Devanagari words hang from one continuous line, so a tilted photo tilts every word along it. Tesseract’s own guidance is that line segmentation degrades when a page is skewed, so rotate the page until the headlines are horizontal before recognising.

• Do not crop the matras above the line

  Reph (र्), ं, ँ, े, ै, ो and ौ sit above the headline. Leave a clear margin above the first line (Tesseract’s docs suggest about ten pixels) so the top row of signs is not cut off or hidden by a shadow.

• Light the lower marks as well

  ु, ू, ृ and the half-forms at the foot of conjuncts (such as द्य or ट्ट) are small and sit under the line. Photograph with even light and avoid a phone’s shadow falling across the bottom of the text.

• Mind the size

  On our test pages, text drawn 18 pixels tall was read without a single character error, while 13 pixel text lost a vowel sign and a digit. Move closer or crop to the text instead of photographing a whole desk.

• Check page segmentation if a page returns nothing

  A Tesseract issue (#1273) describes automatic page segmentation failing on Devanagari pages with the fast model. If a clean page returns empty or partial text, try a tighter crop that contains one column of text.

General advice on resolution, binarisation, skew and borders is in the Tesseract documentation.

Legacy-font PDFs

Hindi PDFs that need the PDF tool, not OCR.

Many Hindi PDFs that look like normal Hindi were typed in a legacy font such as Kruti Dev or DevLys. Their text layer is Latin letters, so reading it directly and decoding the font is exact, while OCR has to guess every sign from pixels. Use the PDF tool first for PDFs in these fonts.

Use the PDF to Unicode tool when the PDF has a text layer. Use OCR for a scan or a photo, which has no text layer, and when the PDF tool says it cannot recognise the font.

5 Hindi legacy fonts the engine converts

• Kruti Dev 010
• Walkman Chanakya 905
• DevLys 010
• DV-TTGanesh
• DVW-TTGanesh

Model download

The Hindi recognition file: about 1.1 MB, downloaded on first use.

Recognition runs in your browser. The first time you press Recognise, the Hindi recognition file downloads from a public file host (jsDelivr); your image is never sent there. The English file, for comparison, is about 4.1 MB. Your browser normally keeps the file, so later visits on the same browser skip the download; clearing site data removes it.