Tools,

Punjabi OCR: nine free Tesseract runs, tested

85.9%

of letters, with pan best and one flag

Get the command

Page mode beats a model swap

We ran Tesseract 5.5.4 on 28 proofread Gurmukhi pages from 10 books on pa.wikisource.

28 pages: letters and whole words right
Run Letters Words
pan best, --psm 6 85.9% 77.2%
pan, --psm 6 85.0% 75.0%
pan, tessdata repo 84.4% 76.3%
pan best 84.4% 76.1%
pan, Sauvola threshold 84.2% 74.6%
script/Gurmukhi best 84.0% 75.4%
pan 83.5% 73.8%
script/Gurmukhi 83.2% 72.7%
pan+eng 79.2% 71.7%

The five Punjabi models sit within 1.2 points. The page mode moved more: --psm 6 added 1.6 points to pan best.

Without the 2018 book of phone photos, pan best with --psm 6 read 93.7%.

Letters are proofread letters found in order. Words are whole words right.

The command: one model, one flag

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

It writes page.txt. The model file is 11.9 MB.

--psm 6 tells Tesseract to read the page as one block of text. The default, --psm 3, finds the blocks itself.

With the flag, pan best took a median 4.5 s a page. Without it, 5.1 s.

Short on time? The fast pan with --psm 6 read 85.0% at 2.0 s a page. Homebrew's brew install tesseract-lang has that fast pan, not pan best.

Orange paper gains the most

Each book, pan best: letters found
Printed Default --psm 6
1964 96.0% 96.4%
2010 82.7% 88.8%
2018, photos 54.3% 51.1%
No year given 94.8% 95.0%
No year given 97.4% 97.4%
No year given 93.0% 93.0%
No year given, orange paper 79.6% 94.4%
No year given 89.0% 88.9%
No year given 94.7% 93.9%
No year given 88.3% 92.8%

The default mode skipped whole lines on the orange paper and in the 2010 poems. The flag got most back.

The 2018 book is phone photos, and its lines curve near the spine. Every run read 55% or less of it.

Page by page, the flag read more on 14 pages and fewer on 6.

The top slips: tippi and bindi

28 pages: pan best with --psm 6, top slips
Printed Read as Times
ਂ bindi Nothing 61
ੰ tippi ੈ vowel sign ai 40
ੰ tippi ੌ vowel sign au 19
ਾ vowel sign aa Nothing 18

The 28 pages hold 540 tippi and 851 bindi.

pan best with the flag wrote 373 tippi and 685 bindi.

It wrote ਸੈਗ 18 times and the word ਮੈ 15 times. The books hold neither. They have ਸੰਗ and ਮੈਂ.

A scanned PDF: use the same flag

mkdir -p tessdata
cp -r "$(brew --prefix)/share/tessdata/configs" tessdata/
curl -L -o tessdata/pan.traineddata \
  https://raw.githubusercontent.com/tesseract-ocr/tessdata_best/4.1.0/pan.traineddata
TESSDATA_PREFIX=$PWD/tessdata ocrmypdf -l pan --jobs 2 \
  --tesseract-pagesegmode 6 \
  --sidecar book.txt scan.pdf book.pdf

On our 28-page PDF: 107.5 s.

book.txt held 84.5% of letters, 79.2% without the flag.

The same text, read back four ways
From Letters in order Words
book.txt 84.5% 76.1%
book.pdf, pdftotext 82.1% 65.0%
book.pdf, pdftotext -raw 84.5% 63.6%
book.pdf, PDFKit (macOS) 83.9% 75.1%

PDFKit, the Mac framework behind Preview, read 704 ਿ signs from book.pdf. book.txt has 785. Search book.txt.

The PNG files read 85.9%. We did not test why the PDF read less.

Without the cp line (Homebrew), OCRmyPDF stops with TesseractConfigError.

Four more things we checked

How we ran it
  • Pages marked proofread or validated on pa.wikisource, from a seeded shuffle of books, 3 a book (2 for two books). Pages with footnotes, tables or running heads in the text were left out.
  • All 28 pages came from PDFs, drawn at most 1,920 px wide.
  • Scores count Gurmukhi letters and signs. Digits, punctuation, Latin letters and spaces are left out.
  • Letters found in order is the longest common subsequence. Slips 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, 2 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 28 scans as JPEGs, one a page.
  • Licenses: Tesseract and tessdata are Apache-2.0, OCRmyPDF is MPL-2.0. Of the 10 scan files, 6 are on Commons (5 CC BY-SA 4.0, 1 public domain) and 4 on pa.wikisource (3 PD-India, 1 untagged). We show no page images.

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 Punjabi.

Try the Actor

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

Get one page free

Questions people ask

Does Tesseract support Punjabi?

Yes, in Gurmukhi, with pan or script/Gurmukhi. On our 28 pages, pan read 83.5% of letters, and pan best with --psm 6 read 85.9%.

Which Tesseract model is best for Punjabi?

The five Punjabi models sit within 1.2 points: 83.2% to 84.4% of letters. The page mode matters more than the model.

Why does Tesseract skip lines?

In the default page mode it lost whole lines on orange paper and in a book of poems. With --psm 6, pan best read 85.9%, not 84.4%. The flag did not help curved phone photos.

Can I OCR a Punjabi PDF?

Yes. OCRmyPDF with -l pan (pan best) and --tesseract-pagesegmode 6 found 84.5% of letters on our 28-page PDF. Without the flag it found 79.2%.

Can Tesseract read Shahmukhi Punjabi?

Not with pan: its character set has no Arabic letters. tessdata has no Shahmukhi Punjabi model, and we did not test urd on Shahmukhi.