Tamil OCR — image to text

Tamil · measured on this model

What the Tamil model got wrong in our test.

Tamil has no headline and no conjunct clusters; instead it relies on the pulli (the dot that silences a consonant) and on vowel signs that sit before, after or around a letter. On our test pages the Tamil model read clean and moderately small text with very few character errors, then failed badly on a blurred, tilted photo. It also does two things that are invisible but matter if you search or compare the text.

How we tested: we wrote 12 lines of Tamil (367 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 Tamil 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

  0.5% character errors

  Engine confidence 95 · 27 invisible U+200C added

• Clean render, serif, 36 px

  0.5% character errors

  Engine confidence 95 · 27 invisible U+200C added

• Small text, sans, 18 px

  0.5% character errors

  Engine confidence 94 · 27 invisible U+200C added

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

  0.5% character errors

  Engine confidence 95 · 27 invisible U+200C added

• Tiny text, sans, 13 px

  2.2% character errors

  Engine confidence 91 · 27 invisible U+200C added

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

  36.5% character errors

  Engine confidence 87 · 4 of 12 lines omitted · 17 invisible U+200C added

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

  1.4% character errors

  Engine confidence 93 · 27 invisible U+200C added

Proofreading checklist

Tamil: 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. ‌ ZWNJ

   Observed · 7 of 7 conditions

   01Invisible joiner characters after pulli letters

   Our Tamil model puts an invisible U+200C ZERO WIDTH NON-JOINER after many words that end in a pulli letter (for example after க் and ம்). You cannot see it, but it sits between the pulli and the space and can break a search, a spreadsheet match or a spell-check. If you need clean text, find and replace U+200C with nothing; it does not change how the word looks.

2. கௌ கெள

   Observed · 7 of 7 conditions

   02The two-part vowel sign ௌ

   கௌதமன் came back as கெளதமன் in all seven conditions. The model wrote the vowel as ெ plus the letter ள, which looks the same in many fonts but is a different spelling in Unicode (ௌ is U+0BCC, which decomposes to U+0BC6 + U+0BD7, not to ள). Search the output for ெள.

   • கௌதமன் → was read as கெளதமன்

3. .

   Observed · 2 of 7 conditions

   03Added full stops

   At 13 pixels and in faded print the model added a full stop after some words (தொழில்நுட்பம்., 6789.). Delete full stops that are not in the source.

4. 6789

   Observed · 1 of 7 conditions

   04Latin digits in a phone number

   In the worst photo, the number 6789 was read as 670. Numbers such as phone numbers, dates and amounts must be compared digit by digit.

   • 6789 → was read as 670

5. (4 of 12 lines)

   Observed · 1 of 7 conditions

   05Heavy blur and tilt

   In our worst condition (22 pixels, 3 degrees tilt, heavy blur and compression) four of twelve lines were omitted, the character error rate was 36.5% and the engine’s confidence still showed 87. The small dots of the script disappear first. If a photo looks soft, retake it before recognising.

   • மாணவர்களுக்கான → was read as ண்வர்களுக்கான

6. ன ண ந ல ள ழ ர ற

   Look-alikes · not measured

   06Look-alike letter families

   Tamil has several families of similar rounded letters. We did not measure them separately, so treat them as a proofreading habit: re-read names and places that contain ன/ண/ந, ல/ள/ழ or ர/ற, and check that a pulli (்) was not read as a full stop or lost.

Photographing Tamil

Tips for photographing Tamil documents.

• Light the dots

  The pulli and the small marks that separate similar letters are a few pixels across. Hold the camera square to the page, use even light, and avoid a glossy page reflecting a lamp.

• Avoid tilt and blur together

  In our test the model coped with 1.5 degrees of tilt and light blur with no more errors than a clean render, but combining 3 degrees of tilt with heavy blur and compression lost a third of the characters. Steady the phone and rotate the image before recognising.

• Vowel signs on both sides

  ெ, ே and ை are written before the consonant and ொ, ோ and ௌ surround it. Leave a margin on the left and right of each line so those signs are not cropped, and check the order in the result.

• Minimum size

  The model read text 18 pixels tall about as well as 36 pixel text in our test, and made roughly four times more character errors at 13 pixels. Crop closer rather than reducing the page.

• Large model, patient first run

  The Tamil model is the largest of the four on this page, so the first recognition downloads more data than Hindi or Bengali.

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

Legacy-font PDFs

Tamil PDFs that need the PDF tool, not OCR.

Tamil publishing in legacy fonts is unusually diverse: Bamini, TSCII, TAB and TAM, and many newspaper fonts each have their own encoding. A PDF set in one of them has a Latin or symbol text layer that the PDF tool can decode when the font is recognised, which avoids the dropped dots, vowel-sign mistakes and invisible characters of OCR.

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.

20 Tamil legacy fonts the engine converts

• Bamini
• Anjal
• Boomi
• Dinamani
• Dinathanthy
• Kavipriya
• Murasoli
• Nakkeeran
• TAB
• TAM / TAM-Elango
• TSCII
• Pallavar
• and 8 more on the Tamil hub

Model download

The Tamil recognition file: about 3.2 MB, downloaded on first use.

Recognition runs in your browser. The first time you press Recognise, the Tamil 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.