Sumatra PDF Plus: a Windows reader fork for Chinese ebooks, OCR and TTS

What Sumatra PDF Plus adds to upstream SumatraPDF

Upstream SumatraPDF is a small Windows reader for PDF and a long list of other formats. This fork keeps that base and targets a narrower reader: someone working through Chinese ebooks and scanned documents on a Windows machine, often offline. The README describes the project as an unofficial fork of SumatraPDF and states it is not affiliated with sumatrapdfreader.org. Product name is Sumatra PDF Plus, and the executable is SumatraPDF-Plus.exe.

The additions listed in the README are OCR, smart PDF dark mode, read-aloud with word-by-word highlighting, an offline dictionary, Chinese EPUB/MOBI handling, and a selection toolbar with highlight, underline, strike out and an Ask AI action. Supported formats include PDF, EPUB, MOBI, AZW/AZW3, FB2, CHM, CBZ/CBR, DjVu and XPS. Scanned PDF, DjVu, images and comics can be recognized locally. If you only read English PDFs and never touch a scanned page, most of this fork is weight you will not use.

How the OCR, dictionary and TTS pipeline is wired

The design is folder-based rather than service-based. OCR runs offline with RapidOCR / PP-OCR Chinese mobile models in {exe}\ocr\, which the README says must contain onnxruntime.dll and the .onnx models. Dictionary lookup reads .idx and .dat files from {exe}\dict\. Traditional Chinese lookup additionally needs the opencc folder. Fonts live in a fonts folder beside the executable, and the quick start warns against copying only the exe.

The data flow for OCR is: page image goes to the local model, the recognized text is held in memory for the current session, and the reader then treats that text as selectable, searchable, look-up-able and readable aloud. Recognize all pages re-OCRs every page even when a text layer already exists, which the README frames as a fix for garbled dual-layer PDFs. If the document has no outline, bookmarks are extracted after OCR. Read-aloud then walks the extracted text with word-by-word highlighting, at speeds from 0.25x to 2.0x. The dictionary is the last step: enable the toolbar button and double-click a word.

The weak point is that these three features depend on each other in one direction. TTS and lookup operate on text the reader can extract, so a scanned page with no text layer produces nothing useful until OCR has run. The README states this order explicitly, with the speaker icon described as coming after OCR.

Installing Sumatra PDF Plus on Windows 10 or 11

There is no installer described. The README gives a portable workflow: extract the whole folder to any path, avoiding special characters in the path, then run the executable. The releases page at github.com/dengxibo/sumatrapdf-plus/releases is where the builds are published; the most recent listed release is 3.7.31 from 2026-09-14.

The one rule that matters is that the folder structure must survive extraction. The README lists the folders that must stay beside the executable: fonts, ocr with onnxruntime.dll and the .onnx models, opencc for Traditional Chinese lookup, and dict holding the .idx and .dat dictionary files.

Start the reader, then toggle the theme and the word-lookup button from the toolbar. For a first real use, open a scanned PDF and click Auto OCR, which the README says is off by default. The dropdown beside it offers Recognize all pages, Recognize all pages and save, and OCR region. OCR region is the safe one to try first: drag a rectangle and the text is copied to the clipboard, with nothing written back to the file.

Read-aloud is started from the speaker icon in the toolbar, or from the Read Aloud (TTS) entry in the menu bar, which offers start from top, cursor or selection. The README notes that default Windows SAPI voices such as Huihui, Kangkang and Yaoyao sound robotic, and points to NaturalVoiceSAPIAdapter for natural speech.

The destructive save path and other limits worth knowing

Recognize all pages and save shows progress and then overwrites the current PDF. The README is direct about this: no Save As, PDF only, and it prompts before replacing an existing text layer or outline. A prompt is not a backup. If the recognition output is worse than the original layer, the original is gone unless you kept a copy. The README does not document rollback or an undo for this operation, and it does not describe a side-by-side output file. Treat that command as a write to your only copy.

Cancel stops queued jobs and keeps pages already recognized, which is reasonable, but it does not undo the pages that finished.

The platform boundary is the bigger one. This is a Windows document reader; every feature described depends on Windows folders, Windows SAPI voices and a Windows executable. The related searches show people looking for an Android version, and the README describes none. There is no macOS or Linux build mentioned either. If your reading happens on a phone or a Mac, this project is the wrong tool regardless of how well the OCR works.

Two more constraints from the README: the OCR models are Chinese mobile models, so recognition quality on other scripts is not something the documentation addresses, and the natural-voice path requires installing a separate project rather than shipping with better voices.

Sumatra PDF Plus versus upstream SumatraPDF and other readers

The honest alternative is upstream SumatraPDF itself. It is the same rendering core and the same format list, and it is the project this fork tracks. The difference in approach is scope: upstream concentrates on being a fast, small reader, while this fork adds a recognition and speech layer plus Chinese-specific handling for EPUB and MOBI. If your documents already carry a good text layer and you read in a language whose SAPI voices you are happy with, upstream gives you the same reading experience without the extra folders.

A second alternative is a full ebook manager with a built-in viewer. Those usually maintain a library, metadata and sync, and they typically do not run OCR locally on a scanned PDF page by page. The trade is the other way around: you get organization and lose the offline recognition pipeline.

What distinguishes this fork is that the OCR, the dictionary and the speech all stay on the machine. The README states the OCR runs offline, and the dictionary is file-based in {exe}\dict\. For anyone reading sensitive or simply large scanned documents, that is the reason to pick it over a cloud-backed reader, and it is also why the folder layout is not optional.

Licence, maintenance and what an upgrade actually costs

The licence is GPL-3.0, with COPYING and AUTHORS in the repository and a COPYING.BSD file also present at the top level. The README states the repository exists for GPLv3 compliance, meaning the source is published to satisfy the source-accompanying-distribution requirement that comes with distributing a GPLv3 binary. That is a statement about obligation, not about support. If you redistribute this build inside an organization, the GPLv3 terms travel with it; that is a question for your own legal review, not something the README settles.

The repository is not archived, and the last push was on 2026-09-15, with releases 3.7.29, 3.7.30 and 3.7.31 landing between 2026-08-23 and 2026-09-14. The version numbers track upstream SumatraPDF, so upgrades arrive as new builds rather than as an in-app updater in the README.

The upgrade cost is the folder model. Because OCR, dictionary, fonts and opencc live beside the executable, replacing the exe alone is not an upgrade path; the README's quick start warns against copying only the exe. Each new release means re-extracting the full folder and re-checking that your dict files and any custom settings survived. Advanced settings named in the README, AutoOcrScanPages, ReadAloudVoiceId and ReadAloudSpeakingRate, are the ones most likely to need re-checking after a version jump.

Editorial conclusion

Adopt Sumatra PDF Plus if you read Chinese EPUB or MOBI files on Windows and want OCR, read-aloud and word lookup without sending pages to a server. Do not adopt it if you need macOS, Linux or Android, or if you cannot accept that Recognize all pages and save overwrites the current PDF in place. Before relying on it, verify that the fonts, ocr, dict and opencc folders sit next to SumatraPDF-Plus.exe, and test the save path on a copy of one scanned PDF.

Frequently asked questions

Official sources

1. dengxibo/sumatrapdf-plus on GitHub
2. License: GPL-3.0
3. Project website
4. README
5. Releases