Kannada OCR: six free Tesseract runs, tested
94.8%
of letters, with Tesseract's fast kan
Five Kannada runs, half a point apart
We ran Tesseract 5.5.4 on 29 proofread pages from 10 printed books on kn.wikisource.
| Model | Letters | Words |
|---|---|---|
| kan | 94.8% | 73.5% |
| kan, tessdata repo | 94.8% | 73.5% |
| kan best | 94.7% | 73.8% |
| script/Kannada best | 94.7% | 72.0% |
| script/Kannada | 94.2% | 69.4% |
| kan+eng | 91.0% | 72.9% |
The fast kan took a median 2.4 s a page. kan best took 3.2 s.
The tessdata repo's kan is the fast network plus a config file. It wrote the same text on all 29 pages.
Letters are proofread letters found in order. Words are whole words right.
The command: one model file
curl -LO https://raw.githubusercontent.com/tesseract-ocr/tessdata_fast/4.1.0/kan.traineddata
mkdir -p tessdata && mv kan.traineddata tessdata/
tesseract page.png page --tessdata-dir tessdata -l kan
It writes page.txt. The model file is 3.6 MB.
Homebrew's brew install tesseract-lang has this same
fast kan. Then -l kan is enough.
The book matters more than the model
Across books, the fast kan moved 11 points: from 87.8% to 98.8% of letters.
| Printed | Letters | Words |
|---|---|---|
| 1913 | 87.8% | 46.7% |
| 1915 | 88.1% | 42.8% |
| 1916 | 96.6% | 79.0% |
| 1917 | 98.7% | 83.6% |
| 1930 | 94.7% | 71.3% |
| 1933 | 96.0% | 80.5% |
| 1967 | 94.6% | 74.1% |
| 1997 | 97.7% | 83.8% |
| 2015 | 98.8% | 91.7% |
| No year given | 92.9% | 68.2% |
The 1913 and 1915 books read worst. Age is not the whole story: the 1917 book read 98.7%.
Our Kannada words average 7.1 letters and signs, and one wrong letter spoils a word.
The top slip: a lost anusvara
We lined up each page's letters with the proofread text and counted one-letter slips.
| Printed | Read as | Times |
|---|---|---|
| ಂ anusvara | Nothing | 67 |
| ಿ vowel sign i | Nothing | 31 |
| ವ | ನ | 25 |
| ನ | ಸ | 20 |
| ಶ | ಕ | 19 |
kan best makes the same five slips, 18 to 65 times each. The bigger model does not fix them.
A scanned PDF: search the sidecar
mkdir -p tessdata
cp -r "$(brew --prefix)/share/tessdata/configs" tessdata/
curl -L -o tessdata/kan.traineddata \
https://raw.githubusercontent.com/tesseract-ocr/tessdata_fast/4.1.0/kan.traineddata
TESSDATA_PREFIX=$PWD/tessdata ocrmypdf -l kan --jobs 2 \
--sidecar book.txt scan.pdf book.pdf
On a 29-page PDF of the same scans it took 57 s. book.txt held 93.8% of letters.
| From | Letters in order | Words |
|---|---|---|
| book.txt | 93.8% | 72.2% |
| book.pdf, pdftotext | 89.6% | 52.4% |
| book.pdf, pdftotext -raw | 93.8% | 54.7% |
| book.pdf, PDFKit (macOS) | 94.1% | 72.2% |
pdftotext -raw puts spaces inside words: 5,091 spaces, where book.txt has 4,244.
PDFKit, the Mac framework behind Preview, kept the words too. Without
the cp line (Homebrew), OCRmyPDF stops with
TesseractConfigError.
Three more things we checked
- kan+eng wrote 962 Latin letters on 28 pages that have none. Leave off +eng: kan and kan best wrote none.
- No Kannada model here can write ೞ or ಱ. Our 29 pages hold neither.
- Apple Vision on macOS 26.6 lists 30 languages, not Kannada.
How we ran it
- Pages marked proofread or validated on kn.wikisource, from a seeded shuffle of books, 3 a book (2 for one book). Pages with footnotes, tables or running heads in the text were left out.
- 12 pages came from PDFs and 17 from DjVu files, drawn at most 1,920 px wide.
- Scores count Kannada 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 29 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: 7 public domain, 3 CC BY-SA 4.0. 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 Kannada.
Or send us the scan. Tack, our AI, does the job for $29, up to 50 pages. First page free. For Kannada, ask first.
Questions people ask
Does Tesseract support Kannada?
Yes, with kan or script/Kannada. On our 29 pages, the fast kan read 94.8% of letters, kan best 94.7%.
Which Tesseract model is best for Kannada?
kan and kan best tie: 94.8% and 94.7% of letters, 73.5% and 73.8% of words. The fast kan is quicker, a median 2.4 s a page against 3.2 s.
Is there a free Kannada OCR?
Tesseract and its models are free (Apache-2.0) and run offline. OCRmyPDF (MPL-2.0) adds a text layer to a scanned PDF.
Can I OCR a Kannada PDF?
Yes. OCRmyPDF with -l kan found 93.8% of letters on our 29-page PDF. Its --sidecar text had 72.2% of words right, pdftotext 54.7% at most.
Should I add English, as kan+eng?
Only if you need the English words. kan+eng read 91.0% of letters, kan 94.8%, and kan+eng wrote 962 Latin letters on 28 pages that have none.