DocTable
How accurate is DocTable's OCR? 97.5% on a synthetic test set
Updated October 2026
On DocTable's own test set of 12 synthetic documents — invoices, bank statements and receipts, including 2 deliberately degraded copies and 2 Arabic documents — 508 of 521 values (97.5%) were read exactly in October 2026. Invoices scored 160 of 160, bank statements 303 of 315 and receipts 45 of 46. These are clean, typeset, single-page documents with known values; real phone photos with blur or glare, handwriting and dense forms score lower, so review results before relying on them. Your first 3 pages are free — scan and download, no card needed. After that, $3 buys 20 pages (one-time, no subscription). Try your first page without an account — sign in only to download it.
| Documents | Values exact | Score |
|---|---|---|
| Invoices | 160 / 160 | 100% |
| Bank statements | 303 / 315 | 96.2% |
| Receipts | 45 / 46 | 97.8% |
| — of which degraded copies | 33 / 33 · 65 / 65 | 100% |
| — of which Arabic invoice | 28 / 28 | 100% |
| — of which Arabic statement | 43 / 55 | 78.2% |
| All 12 documents | 508 / 521 | 97.5% |
How it works
- Upload a document of your own — PDF, scan or phone photo.
- Compare the extracted rows with the page in the preview.
- Rows whose arithmetic does not add up are flagged for you to check.
- Download Excel, CSV or Word — included in your 3 free pages.
How it was measured
The 12 documents were generated by a script from data written into it, so every value on every page is known in advance: 4 invoices (1 in Arabic), 4 bank statements (1 in Arabic, with Arabic-Indic digits), 2 receipts, and 2 deliberately degraded copies — one rotated, shadowed, blurred and saved as a low-quality JPEG, the other dimmed, low-contrast and rotated. No customer data was used. Each document is a single page and went once through the same extraction DocTable runs for users.
Every header field and every table cell was compared with the known value as an exact match, after normalising whitespace, thousands separators and Arabic/Persian digits. A value that is close but not identical counts as a miss.
What it got wrong
13 of 521 values were missed:
- 1 on a receipt — the store name was read in the text but not returned as its own field.
- 12 on the Arabic bank statement — the account number and the statement period, printed right to left, came back with their digit groups in the order they appear on screen rather than the order they were written; and all 10 dates came back as day-month-year (03-10-2026) where the truth was year-month-day (2026-10-03). The dates are the same days in a different order, but they are counted as misses, because a spreadsheet import expecting year-month-day would read them wrongly.
What the number does not cover
The set is 12 synthetic, cleanly typeset, single-page documents. Only 2 are degraded and only 2 are Arabic; none is handwritten, and none is a real phone photo with perspective or glare. Totals were not cross-checked in this run. Dense forms such as a Persian customs declaration are a harder case and are not in this number. A real-world figure will be lower than 97.5%.
How mistakes are caught
Because no OCR is exact on every page, DocTable checks the arithmetic of what it reads: running balances on statements, quantity × unit price, line items against the subtotal, and subtotal plus tax against the total. Rows that do not add up are flagged in the preview and in the Excel file, so you check those rows instead of every cell. The check covers numbers that should add up; dates, names and descriptions still need your eye.
Test it on your own document
A published figure tells you how a method performs on its own test set; your documents are the test that matters. Upload one and compare the preview with the page. Related guides: bank statement to Excel, invoice to Excel, Arabic OCR and how the AI OCR works.
Frequently asked questions
How accurate is OCR?
It depends on the document: on DocTable's synthetic test set of 12 clean, typeset documents, 508 of 521 values (97.5%) were read exactly. Real phone photos, handwriting and dense forms score lower, so treat any single figure as a ceiling and check the result.
Is there an OCR accuracy benchmark for invoices and bank statements?
Yes, DocTable publishes its own: invoices 160 of 160 values, bank statements 303 of 315 and receipts 45 of 46, on synthetic single-page documents in October 2026. It is not an independent benchmark, and the method and misses are on this page.
How was DocTable's accuracy measured?
12 documents were generated from known data, each run once through DocTable's normal extraction, and every header field and table cell was compared with the known value by exact match after normalising whitespace, thousands separators and Arabic/Persian digits.
How accurate is handwriting OCR?
Not measured: DocTable's test set has no handwritten documents, so no handwriting figure is published. Neat print reads best, and cursive, crossed-out or squeezed entries are where errors appear — see handwritten table to Excel.
Which is the most accurate PDF to Excel converter?
No independent comparison of PDF to Excel converters is published, so no honest answer names one. DocTable's own figure is 97.5% on the synthetic set described here; the quicker test is your own document — your first 3 pages are free, download included.
How does DocTable catch OCR mistakes?
It checks the arithmetic it reads — running balances, quantity × unit price, line items against the subtotal, subtotal plus tax against the total — and flags rows that do not add up in the preview and in the Excel file.
Related pages
Figures from DocTable's own synthetic test set (October 2026), not an independent benchmark. Always verify OCR output before financial or legal use.