Telugu OCR: six free Tesseract runs, tested
90.5%
of letters, with Tesseract's script/Telugu best
Five models, one point apart
We ran Tesseract 5.5.4 on 29 proofread pages from 10 printed books on te.wikisource.
| Model | Letters | Words |
|---|---|---|
| script/Telugu best | 90.5% | 51.9% |
| tel best | 90.3% | 51.6% |
| tel, tessdata repo | 90.3% | 51.3% |
| tel | 89.9% | 51.3% |
| script/Telugu | 89.4% | 49.5% |
| tel+eng | 80.5% | 48.9% |
The five Telugu models sit within 1.1 points of each other.
The fast tel took a median 1.9 s a page. script/Telugu best took 3.2 s.
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_best/4.1.0/script/Telugu.traineddata
mkdir -p tessdata && mv Telugu.traineddata tessdata/
tesseract page.png page --tessdata-dir tessdata -l Telugu
It writes page.txt. The model file is 14 MB.
Homebrew's brew install tesseract-lang has the fast
tel. Use -l tel (89.9% here, and quicker).
The book matters more than the model
Between models, the score moved about a point. Between books, it moved 22 points.
| Printed | Letters | Words |
|---|---|---|
| 1891 | 77.4% | 25.8% |
| 1902 | 93.3% | 61.1% |
| 1916 | 86.3% | 33.8% |
| 1929 | 89.4% | 32.9% |
| 1931 | 88.1% | 39.2% |
| 1932 | 87.3% | 37.5% |
| 1934 | 83.6% | 27.6% |
| 1955 | 89.3% | 35.9% |
| 1960 | 95.8% | 68.0% |
| 1988 | 99.7% | 98.1% |
The 1960 and 1988 books came out best. They were also our smallest images, about 1,000 px wide.
Why so few whole words? Our Telugu words average 7.9 letters and signs, and one wrong letter spoils a word.
No official model writes ఁ or ఱ
The character sets of all five Telugu models lack ఁ (arasunna) and ఱ. 7 of our 10 books use ఁ, and 9 use ఱ.
| Letter | In the books | Written by any run |
|---|---|---|
| ఁ | 244 | 0 |
| ఱ | 64 | 0 |
283 of our 3,722 proofread words hold a letter the models lack, so no run gets them right.
| Printed | tel best |
|---|---|
| శివుఁడన | శివుండన |
| కొందఱును | కొందఅును |
tel best wrote 1,233 ం. The books hold 994 ం and 244 ఁ.
Search tip: type ం for ఁ. With tel best, 42 of the 227 words with ఁ match, against 0 without it.
A scanned PDF: OCRmyPDF works
mkdir -p tessdata
cp -r "$(brew --prefix)/share/tessdata/configs" tessdata/
curl -L -o tessdata/tel.traineddata \
https://raw.githubusercontent.com/tesseract-ocr/tessdata_best/4.1.0/tel.traineddata
TESSDATA_PREFIX=$PWD/tessdata ocrmypdf -l tel --jobs 2 \
--sidecar book.txt scan.pdf book.pdf
On a 29-page PDF of the same scans it took 90 s. book.txt held 90.2% of letters.
| From | Letters in order | Words |
|---|---|---|
| book.txt | 90.2% | 51.7% |
| book.pdf, pdftotext | 86.4% | 51.6% |
| book.pdf, pdftotext -raw | 90.2% | 51.5% |
| book.pdf, PDFKit (Preview) | 89.9% | 51.7% |
Plain pdftotext cut lines into pieces, so letters fell out of order.
Add -raw.
The cp line is for Homebrew. We ran it on a Mac. Skip it and OCRmyPDF
stops with TesseractConfigError.
Two more things we checked
- tel+eng wrote 1,862 Latin letters on 24 pages that have none. Leave off +eng: the tel models wrote none.
- Apple Vision on macOS 26.6 lists 30 languages, not Telugu.
How we ran it
- Pages marked proofread or validated on te.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.
- 23 pages came from PDFs, drawn 1,920 px wide. 6 came from two DjVu files, 999 and 1,065 px wide.
- Scores count Telugu letters and signs. Digits, punctuation, Latin letters and spaces are left out.
- Letters found in order is the longest common subsequence. Python's difflib, which our other language pages used, counts fewer, most of all when a page repeats a phrase.
- With difflib, tel best scored one 1891 page 32.6%, not 80.1%. Totals rose by 1.3 points or less.
- 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. The 10 scan files are on Commons: 8 public domain, 2 CC BY-SA 3.0. 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 Telugu yet.
Or send us the scan. Tack, our AI, does the job for $29, up to 50 pages. First page free. For Telugu, ask first.
Questions people ask
Does Tesseract support Telugu?
Yes, with tel or script/Telugu. On our 29 pages, script/Telugu best read 90.5% of letters, the fast tel 89.9%.
Why is ఁ missing from my Telugu OCR text?
None of the five official Tesseract Telugu models has ఁ or ఱ in its character set. Our runs wrote 0 of the 244 ఁ and 64 ఱ in the books.
Is there a free Telugu 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 Telugu PDF?
Yes. OCRmyPDF with -l tel and the tel best model found 90.2% of letters on our 29-page PDF. Read its text with pdftotext -raw, or keep --sidecar.
How accurate is Telugu OCR on old books?
On 9 books printed from 1891 to 1960, script/Telugu best read 89.0% of letters and 42.7% of words. On a 1988 book: 99.7% and 98.1%.