Turn PDF text or image-only pages into an editable DOCX, then review the reconstructed text and tables.
or drop one PDF here
No file selected
Choose a PDF to start.
Editable text and inferred tables. Original images, drawings and exact page design are not exported.
One unencrypted PDF · 30 MB · 60 source pages, or 10 when OCR is needed. Review all limits.
In your browserNo uploadsNo account
If you need to change PDF to DOCX, the result is a text reconstruction rather than an editable replica of the PDF. Source line breaks may become paragraph breaks. Page boundaries are kept with explicit Word page breaks, but Word can reflow the text onto additional pages.
The default mode looks for a repeated wide gutter to read two columns separately. Three or more consistently aligned text starts can become a table.
These are geometric heuristics: a two-column table, sidebars, merged cells or wrapped cell contents can be ambiguous. The preview reports which strategy was used for each page. Simple line order sorts by height and then left to right, making the comparison repeatable with the same input.
Simple stacks of separately positioned CJK characters can be read top to bottom, right column first. They are normalized to horizontal paragraphs; vertical typography is not retained. Rotated text and font-defined vertical writing stop conversion.
Image-only pages automatically use local OCR. English is the default; choose Simplified Chinese + English for simplified Chinese scans. OCR word positions enter the same reconstruction heuristics.
A page with an existing text layer keeps the original extraction path; images within that page are not OCRed.
This PDF to DOCX converter online reconstructs editable text and inferred tables from a PDF with selectable text or an image-only scan.
What is not retained: Original images, charts, fonts, forms, links, annotations and accessibility structure are not exported. Exact page design is not retained.
It stops if a page-processing checkpoint exceeds 90 seconds.
Very complex pages can still make the browser unresponsive.
Text in an image remains unavailable when another part of that page has extractable text.
Rotated pages still stop conversion; no partial DOCX is offered on a failure.
These are observations from our synthetic files, not an accuracy guarantee for other PDFs. Current outputs and remaining errors appear first; settings and earlier behavior are recorded below.
The text-layer comparison was run on 13 October 2026. The image-only scan row includes the OCR update tested on 16 October 2026.
On a small screen, swipe across the table to read every column.
| Input PDF | Observed DOCX output | Remaining limitation |
|---|---|---|
| Image-only scan | Automatic English OCR now produces two editable paragraphs: SCAN-ONLY-001 / This text exists only in page pixels. | This clean fixture matches the source text; it does not establish accuracy on other scans. The original page image is not embedded. |
| Two columns, eight identifiers | LEFT01–LEFT04, then RIGHT01–RIGHT04. All eight identifiers appear once. | Order is reconstructed; the two-column page design is not retained. |
| Three-column table | One native Word table, seven rows including the source header, three cells per row; all values match this fixture. | Merged cells and wrapped cell contents were not tested in this fixture. |
Download the inputs, both DOCX outputs and exact test notes. That historical kit predates OCR: its scan has no output file. The current OCR inputs, outputs and settings are provided below. Repeat these settings with your own PDF and compare every paragraph and table cell.
On a small screen, swipe across the table to read every column.
| Input PDF | Observed DOCX output | Remaining limitation |
|---|---|---|
| English scan, three text lines | SCAN ENGLISH 001 / Invoice total: 123.45 USD / Please review every line. | The final single-column path retains the OCR line text; these three lines match this source. Original design is not retained. |
| English scanned table, four rows × three columns | One native Word table with four rows and three cells per row. Headers: Item / Qty / Price. Quantities: 2, 3, 5. Prices: 12.50, 24.00, 36.50. | PEN001 → PENO001; BOOK002 → BOOKO002; TOTAL003 → TOTALO003. These actual OCR errors remain in the DOCX. Merged cells and wrapped cells were not tested; native cells do not prove character accuracy. |
Download the OCR inputs, actual DOCX outputs and exact check records. Review the uncorrected OCR strings against the source scans. These observations apply to these synthetic files, not a general recognition guarantee.
| Input PDF | Observed DOCX output | Remaining limitation |
|---|---|---|
| Two positioned CJK stacks | Two horizontal paragraphs: 右列自制样本, then 左列保留顺序. | Vertical typography is not retained. This fixture does not cover rotated text or font-defined vertical writing. |
| Input PDF | Observed DOCX output | Remaining limitation |
|---|---|---|
| Simplified Chinese scan, three text lines | 中文扫描测试 / 订单编号 ZH003 / 金额 123.45 元 | Retaining recognized line text for the single-column case removes those spacing errors in this fixture. This is not a test of traditional Chinese, handwriting or other languages. |
We use our own synthetic PDFs and run the same input with both reading-order settings. The comparison below records observed failure points, settings and the resulting DOCX structure. It is a check of these fixtures, not a general accuracy claim or a test of competing tools.
Run date: 13 October 2026. All four sources have one US Letter page (612 × 792 points). Two-column and table text uses Helvetica 11 points; positioned CJK glyphs use a Noto Sans SC fixture font at 18 points. The scan contains a locally generated image only. No customer documents were used.
Comparison settings: select the same PDF, run Simple line order, download, then change to Layout-aware reconstruction and run again. For the text-layer rows, “before” and “after” below mean these two modes in the same converter version, not two historical deployments. The image-only row separately compares the pre-OCR behavior with the OCR update tested on 16 October 2026. No page-size or font setting was changed between runs.
On a small screen, swipe across the table to read every column.
| Input PDF | Earlier behavior / comparison mode |
|---|---|
| Image-only scan | Page 1 has no extractable text; no DOCX produced. |
| Two columns, eight identifiers | LEFT01, RIGHT01, LEFT02, RIGHT02… are interleaved. |
| Three-column table | Header and six records become ordinary paragraphs; no Word table. |
| Two positioned CJK stacks | Left and right characters are mixed on each line. |
We inspected the actual downloaded DOCX packages, not only this page's preview. An independent DOCX parser checked paragraph order and native table cells. Each of the six successful DOCX outputs was reopened with a separate browser DOCX renderer and its rendered page inspected. Word/LibreOffice visual fidelity was not verified: the native LibreOffice renderer was unavailable on the test machine.
These three additional synthetic inputs each contain one image-only US Letter page (612 × 792 points), generated as a 1,836 × 2,376 pixel image with 44-pixel Arial or Microsoft YaHei text. The table has one header and three data rows. All PDFs have no extractable text layer. Settings: Layout-aware reconstruction, English for the English and table samples, Simplified Chinese + English for the Chinese sample; PSM 3, render scale 3, 216 DPI hint. No denoising, thresholding or automatic spelling correction was added.
On a small screen, swipe across the table to read every column.
| Input PDF | Earlier behavior / first OCR prototype |
|---|---|
| English scan, three text lines | The pre-OCR tool rejected the page. In the first OCR prototype, word-box baselines split “Please review every line.” into “Please review line.” and a separate “every”. |
| Simplified Chinese scan, three text lines | The pre-OCR tool rejected the page. The first positional reconstruction introduced spaces into “中文扫描测试” and removed the space before ZH003, although the recognized line text was intact. |
| English scanned table, four rows × three columns | The pre-OCR tool rejected the page. |
We checked the actual browser-downloaded DOCX packages with an independent parser and opened all three new OCR outputs plus the original scan output in LibreOffice 26.8.1. Each rendered as one page; each rendered page was inspected. Microsoft Word itself was not tested. The original columns, table and positioned-CJK fixtures were rerun in both modes: all six DOCX packages have the same unzipped file contents as their pre-OCR baselines. A mixed two-page PDF kept its text-layer page on the original path and OCRed its image-only page. Password-protected and rotated PDFs still stopped without a download; an 11-source-page image-only PDF stopped at the 10-page OCR limit without a partial download.
OCR renders each page at up to 3× scale with a 2,400-pixel longest side, using automatic segmentation (PSM 3) and a 216 DPI hint.
No cold-start or physical-phone timing is claimed here.
The selected PDF is read and the DOCX is created in this browser. Your document and reconstructed text are not uploaded to a conversion server. Conversion code, workers and selected OCR language data load from this site.
The OCR engine and selected language data download from this site only when needed; initial use can take longer and use more memory.
Workers terminate after each conversion; later runs initialize again, while browser resource caching may avoid downloads.
Website-delivery and analytics requests are separate from conversion and are described in our privacy policy; document contents and output are not sent as analytics events.
No. The input must be a PDF, not a standalone photo. To convert scanned PDF to Word, choose one PDF and the OCR language. This scanned PDF to Word workflow supports English and Simplified Chinese. For a scanned document to Word, the text is reconstructed, not an editable facsimile.
Only the existing text layer is used; images on that page are not OCRed or embedded. See reading order and editable output.
The run stops without a partial DOCX. Check Support and limits for the applicable failure conditions.
Page updated: