Tools,

Odia OCR: eight free Tesseract runs, tested

74.2%

of letters: script/Oriya best, --psm 6

Get the command

Use script/Oriya, not ori

We ran Tesseract 5.5.4 on 30 proofread Odia (Oriya) pages from 10 books on or.wikisource.

30 pages: letters and whole words right
Run Letters Words
script/Oriya best, --psm 6 74.2% 32.3%
script/Oriya best, Sauvola threshold 73.8% 32.0%
script/Oriya best 73.6% 32.1%
script/Oriya (fast) 73.5% 31.6%
ori best 71.7% 27.0%
ori (fast) 71.7% 26.8%
ori, tessdata repo 71.7% 26.8%
ori+eng 52.0% 24.4%

The script model beat ori on letters and words. --psm 6 added 158 letters. Sauvola changed little.

tessdata's ori is the fast ori plus one config line. It wrote the same text on all 30 pages.

The command: script/Oriya best, --psm 6

curl -LO https://raw.githubusercontent.com/tesseract-ocr/tessdata_best/4.1.0/script/Oriya.traineddata
mkdir -p tessdata && mv Oriya.traineddata tessdata/
tesseract page.png page --tessdata-dir tessdata -l Oriya --psm 6

It writes page.txt. The 16.3 MB model took a median 3.2 s a page.

Homebrew's brew install tesseract-lang has the fast ori, a 1.5 MB file. It read 71.7% of letters, with about 5 more words in a hundred wrong.

Top slip: the vowel sign i goes missing

30 pages: script/Oriya best, --psm 6, top slips
Printed Read as Times
◌ି vowel sign i Nothing 343
ତ ta ଭ bha 77
◌ି vowel sign i ◌ୁ vowel sign u 63
ନ na ଜ ja 54
◌ି vowel sign i ◌୍ virama 51

The pages hold 2,102 of these i signs.

ori's top slip is a different one: it lost the e sign 288 times.

ori can write ଚ for the e sign

The vowel sign e (◌େ) is drawn before its letter. ori wrote 678 more ଚ than the pages hold, and 830 fewer ◌େ.

1,712 words with ◌େ or ◌ୋ
Run Right ଚ in front instead
ori (fast) 303 123
ori best 312 115
script/Oriya best, --psm 6 599 2

The script model rarely makes this slip.

No run wrote ୱ right

The pages hold ୱ (wa, U+0B71) 90 times, in 89 words. Every run got 0 of those words right.

Where ୱ was printed, script/Oriya best with --psm 6 wrote ଵ (U+0B35) 31 times. The pages hold no ଵ. Every model's character set has ୱ; we did not test why.

Each book: 61.8% to 96.9%

Each book: letters found
Printed script/Oriya best, psm 6 ori
1868, small old type 74.9% 69.1%
1879, dense small type 69.5% 67.3%
1923, dense small type 61.8% 62.4%
1940 73.3% 74.3%
1961, verse, brown paper 69.6% 66.4%
1968, cream paper 75.7% 69.6%
1983 75.1% 70.9%
1984, large type 77.1% 68.8%
2014, clean modern type 96.9% 96.7%
No year, bold type 63.6% 64.0%

Only the 2014 book read well. The 1923 book and the bold book with no year read lowest.

script/Oriya best with --psm 6 beat ori on 7 of 10 books.

A scanned PDF: the same model

mkdir -p tessdata
cp -r "$(brew --prefix)/share/tessdata/configs" tessdata/
curl -L -o tessdata/Oriya.traineddata \
  https://raw.githubusercontent.com/tesseract-ocr/tessdata_best/4.1.0/script/Oriya.traineddata
TESSDATA_PREFIX=$PWD/tessdata ocrmypdf -l Oriya --jobs 3 \
  --tesseract-pagesegmode 6 --sidecar book.txt scan.pdf book.pdf

It took 56.1 s on our 30-page PDF.

The same text, read back five ways
From Letters in order Words
book.txt 74.3% 32.8%
book.pdf, pdftotext 68.6% 26.4%
book.pdf, pdftotext -raw 74.3% 26.4%
book.pdf, pdftotext -layout 73.5% 26.4%
book.pdf, PDFKit (macOS) 74.3% 32.8%

All five readers kept the same 874 i signs and 1,503 e signs as book.txt. pdftotext lost about 6 points of words. Search book.txt.

Without the cp line, OCRmyPDF stopped with TesseractConfigError in our Bengali and Punjabi tests.

Five more things we checked

How we ran it
  • Pages marked proofread or validated on or.wikisource, from a seeded shuffle of books, 3 a book. Pages with footnote tags, tables or templates in the wiki text were left out.
  • The books were printed 1868 to 2014; 1 gives no year. Every page is print. We looked at each one first.
  • Two pages hold 89 proofread letters the scan does not show: an editor's note and a running head. Every run misses almost all of them; without them each run scores 0.1 to 0.3 points higher.
  • All 30 pages came from PDFs, drawn 864 to 1,920 px wide.
  • Scores count Odia letters and signs. Digits, the isshar, the danda, punctuation, Latin letters and spaces are left out.
  • Letters found in order is the longest common subsequence. Words: whole words right. Slips and long skipped stretches (20 or more letters in a row with nothing found) come from a second alignment, so counts are close, not exact.
  • Models from tessdata_fast, tessdata_best and tessdata at tag 4.1.0.
  • One thread a run, 3 runs at a time, at low priority, while other jobs ran. Times are rough.
  • OCRmyPDF 17.13.0, pdftotext from Poppler 26.03.0, PDFKit on macOS 26.6. The PDF holds the 30 scans as JPEGs, one a page.
  • Licenses: Tesseract and tessdata are Apache-2.0, OCRmyPDF is MPL-2.0. All 10 scan files are on Commons under public domain, CC0 or CC BY-SA. We show no page images and quote no passage.

Scanned pages to text, no install

Run it yourself. Our Apify Actor turns a scanned PDF into Markdown or text. $0.006 a scanned page. It reads 18 languages, not Odia.

Try the Actor

Or send us the scan. Tack, our AI, does the job for $29, up to 50 pages. First page free. For Odia, ask first.

Get one page free

Questions people ask

Does Tesseract support Odia?

Yes, with ori or script/Oriya. On our 30 pages, script/Oriya best with --psm 6 read 74.2% of letters and 32.3% of words.

Which Tesseract model is best for Odia?

On our 30 pages, script/Oriya best with --psm 6. It beat ori best by 2.4 points of letters and 5.3 points of words.

Why does Tesseract write ଚ for the Odia e sign?

We did not test why. On our pages, ori did it in 123 words, script/Oriya best with --psm 6 in 2.

Can I OCR an Odia PDF?

Yes. OCRmyPDF with -l Oriya and --tesseract-pagesegmode 6 found 74.3% of letters and 32.8% of words on our 30-page PDF.

Why are the word scores so low?

One wrong letter spoils a word. With script/Oriya best and --psm 6, the 2014 book got 86.7% of words right; the other nine missed 23% to 38% of letters.