doc2docx

doc2docx is a pure-Python converter from Microsoft Word 97–2003 binary .doc files to modern .docx files. The conversion engine follows the published Microsoft Office binary-format specifications and does not invoke Microsoft Word, LibreOffice, COM, Java, or another conversion executable.

The project is usable for a growing set of real documents, but it is not yet a complete implementation of every legacy Word feature. Unsupported or lossy content is reported explicitly instead of being silently presented as a fully faithful conversion.

Highlights

• Reads CFB/OLE Word documents with deterministic, bounded parsers.
• Opens XOR-obfuscated, classic RC4, and RC4 CryptoAPI password-protected documents when a password is supplied.
• Preserves text, common character and paragraph formatting, fonts, styles, native numbered and bulleted lists, tables, page layout, section page and line numbering, paragraph text frames, floating table positioning, headers, and footers.
• Converts footnotes and endnotes—including custom separators, placement, and numbering controls—plus comments, named bookmarks, safe document-local reference fields, common date/metadata/page/statistic fields, and positioned textboxes (including drawing-canvas groups flattened to a filled underlay plus independently positioned children) to native WordprocessingML structures.
• Preserves core document metadata such as title, author, subject, keywords, revision, and creation/modification dates.
• Restores inline and floating PNG, JPEG, BMP/DIB, TIFF, EMF, and WMF pictures in the main document and header/footer stories, including rotation, flips, tight/through wrap polygons (bounding-rectangle contour when OfficeArt omits vertices), empty PictureFrame shapes used only for wrap (no BLIP), and common floating preset shapes such as polygons, diamonds, stars (5/8/16-point), arrows (left/up/down/left-right/up-down, striped, notched, and curved right/left/up/down), smile faces, ribbons, arrow callouts, folded corners, action buttons (blank through navigation/document icons), brackets/braces, explosions/seals, no-symbol/math symbols, chevrons, bevels, waves, lightning bolts, suns, moons, cubes/cans/donuts, flowchart process/decision/data/document and extended flowchart symbols (sort, merge, delay, display, punched tape, etc.), lines, and plaques. Path-only and adjustment-formula presets (including diamonds, flowcharts, classic and modern cans, cubes, donuts, chevrons, pentagons, plaques, and cardinal/left-right arrows) are emitted as Word-style <v:shapetype> references with authoritative VML geometry (including Word SaveAs <v:handles>), and OfficeArt adjustValue/adjust2Value are passed through as VML adj (sparse/short forms are filled from the preset default on emit, e.g. ,14040 → 5400,14040, 8100 → 8100,5400) so lid depth and arrow geometry match the source.
• Preserves confirmed embedded OLE ObjectPool storages as native DOCX embedded objects, and retains Macintosh PICT payloads for consumers that can render them.
• Produces a structured diagnostic report for unsupported, repaired, or approximated source features.
• Writes deterministic OPC packages using only the Python standard library at runtime.

Installation

python -m pip install msdoc2docx

Python 3.10 or newer is required.

Command line

Convert beside the input file:

doc2docx input.doc

Choose the output path and save a JSON report:

doc2docx input.doc -o output.docx --report report.json

For a password-protected document, prefer a UTF-8 password file so the secret does not appear in the process command line:

doc2docx protected.doc --password-file password.txt

Inspect a source file without converting it:

doc2docx inspect input.doc --json

Convert a directory while preserving its relative layout and write one JSON summary for successful and failed files:

doc2docx batch input-directory -o output-directory --recursive --report batch.json

Python API

from doc2docx import convert

result = convert("input.doc", "output.docx", password="secret")
print(result.report.to_dict())

The source document is opened read-only. The destination and JSON reports are written through temporary files and atomically replaced after validation or serialization; the converter will not overwrite the input file. Report paths are also checked against source and output paths, including every input and destination in batch mode.

Current limitations

Grouped/custom OfficeArt geometry and advanced drawing effects remain incomplete. Unsupported geometry with reliable OfficeArt vertices is emitted as an exact VML path in the shape's native coordinate space (including nf/ns from path NoFill/NoLine escapes); geometry with only a wrap contour is retained as an explicitly diagnosed contour approximation; geometry without such evidence is deferred. Linked or unanchored OLE storages are not activated or guessed; only confirmed embedded-object anchors are packaged. Macintosh PICT images are retained as original media, although rendering them depends on support in the DOCX consumer. The legacy macro story has no safe WordprocessingML equivalent and is deliberately omitted. Fields that can execute actions or access external content are kept as cached text. Some legacy layout behavior can only be approximated in WordprocessingML and is called out in the conversion report.

See CHANGELOG.md for milestone and release details.

Related projects

doc2docx follows the same package and command conventions as xls2xlsx and ppt2pptx:

The projects share a consistent API, CLI structure, report format, and exit status conventions. Their conversion engines remain separate because the Word, Excel, and PowerPoint binary formats use different data models.

Development

Run the standard-library test suite from a source checkout:

PYTHONPATH=src python -m unittest discover -v

Build distributable artifacts with:

python -m build

Real-document regression tests may use LibreOffice to create or render test artifacts, but LibreOffice is not used by the converter itself.

Word bilateral visual compare (Windows)

Authoritative page-image regression uses Microsoft Word COM only from scripts/ and optional tests. Install the optional visual stack into a local venv, then compare a source .doc against the converter's .docx:

python -m venv .venv
.venv\Scripts\python -m pip install -r scripts/requirements-visual.txt
$env:PYTHONPATH = (Resolve-Path .\src)
.venv\Scripts\python scripts\word_bilateral_compare.py input.doc -o evidence\case

The tool opens independent read-only temporary copies of the source DOC and output DOCX, exports both to PDF with the same Word instance, rasterizes at a fixed DPI, and writes per-page reference/actual/diff/overlay images plus a manifest.json that includes SHA-256 hashes, page metrics (MAE/RMSE/changed pixel ratio/SSIM), structure counts, and the conversion report diagnostics. LibreOffice rendering remains a development aid only and is not treated as release-level fidelity proof.

Specifications

• MS-DOC: Word Binary File Format
• MS-ODRAW: Office Drawing Binary File Format