OCR Accuracy Comparison: Tesseract vs 6 Other Tools on 6 Scripts

Each cell is the share of characters the tool got wrong on one scanned book page, so lower is better. One page per script is a snapshot, and the reference text itself is a volunteer transcription, which is why gaps under three points are called ties.

FastOCR ranged from 1.0% to 9.1% across the six pages shown (lower is better). On Arabic, Urdu, Bengali, Chinese and English the best other tool was within the noise of FastOCR, so those pages count as ties.

Free web forms behaved very differently from one another. imagetotext.info stayed close to FastOCR on the three pages we ran. PDF24 and Tesseract struggled on Urdu. OnlineOCR.net returned Latin letters for the Urdu page (99.6% wrong, lower is better), because its language list does not include Urdu.

The OCR.space row shows two engines on purpose. The default engine rejected Arabic, while the second engine read it, so on that site the engine menu matters more than the tool. See FastOCR and OCR.space. Smallpdf gave one Arabic result and then stopped offering free downloads.

Character error rate counts the edits needed to turn the OCR output into the correct text, then divides by the length of the correct text. A wrong, missing or extra character is one edit. Word error rate counts whole words instead, so a single wrong letter spoils the word.

Accuracy is normally 100% minus the error rate, so a 5% error rate (lower is better) becomes a 95% accuracy figure. Vendors do not always say which unit they mean, and a page-level or word-level figure can sit far below a character-level one for the same output.

Our scores remove spaces, punctuation and Arabic diacritics from both texts before comparing, so a missing tashkeel mark does not count against a tool.

Tesseract is free, open source and runs on your own machine. We ran version 5.5.4 locally with the matching language pack, one page per script. Its best pages were Tamil and English at 8.5% of characters wrong, and its worst were Chinese, Urdu and Bengali (lower is better; the table above has every figure).

On Arabic, Urdu, Bengali, Tamil, Chinese and English FastOCR was clearly lower. The score is not a verdict on the engine: Tesseract's image-quality guide says resolution, noise, skew and binarisation all change recognition, and a tuned pipeline can beat one untuned run.

Our pages are old, sometimes stained book scans. A clean modern page in a Latin script is the case Tesseract was built around, and we did not measure it. You also choose the language pack yourself. The wrong pack, or none, gives worse output than the figures here, while FastOCR detects the language on its own. For the longer comparison see FastOCR and Tesseract, and the engine's own documentation for the languages it ships.

We did not time any tool, and we hold no dated FastOCR timing to cite, so this page ranks nothing on speed. That includes Tesseract speed. For a hosted tool the wait also includes upload, queueing, page count and server load, which a vendor figure rarely matches.

For Tesseract the main levers are on your side. Its FAQ says version 4 can use the fast integer models and up to four CPU threads per page, that four threads slow things down on a two-core machine, and that OMP_THREAD_LIMIT sets the thread count. Speed is therefore your hardware and settings, not a property of the engine alone.

To get a number you can trust, take ten of your own pages, run each tool on the same files on the same machine or connection, time from submitting the file to holding the text, and repeat at another time of day. Record accuracy for the same run: a quick result that needs a long proofread is slow.

No threshold applies everywhere, because the page decides most of the result. On our six pages FastOCR ranged from 1.0% to 9.1% of characters wrong and Tesseract from 8.5% to 71.9% (lower is better). The same engine can read one script well and another badly.

Turn the rate into wrong characters before judging it: 3% on a 2,000-character page is about 60 wrong characters. That may be fine for search and unacceptable for a price list, a contract or a name.

Scan quality moves the figure as much as the engine. Skew, stains, small print, low resolution and phone shadows all raise it. Treat a single percentage from a vendor with suspicion until you know the script, the kind of page and whether it counts characters or words.

We tested no paid enterprise product, so we cannot say how Tesseract compares with one. Published studies exist, but each covers a narrow document set.

A 2022 paper in the International Journal of Electrical and Computer Engineering (the paper) compared Tesseract with a commercial cloud OCR service on Thai vehicle registration certificates. One document type in one script says little about books or forms. The EffOCR paper (Efficient OCR for Building a Diverse Digital History) reports character error rate against the same kind of service on historical documents, with Tesseract as a baseline, and historical print behaves differently from modern scans.

Read the tables in the papers themselves, and check the document type, language, engine version and date, because engines change. A neutral explainer of the two engines is in the two engines side by side.

Pick ten to twenty pages from your real work, type the correct text for a few lines of each, run every candidate on the same files, and compare character by character. The step-by-step version is in how to benchmark OCR on your own documents.