PDF to Audiobook on GitHub: Local and Offline Options

For a local PDF-to-audiobook workflow, start with ebook2audiobook if you want a documented file-to-audio command, or Abogen if you prefer a desktop interface with chapter controls and subtitles. Both require text you can check, suitable voice models and a working local installation. Downloading an open-source project is only the first step toward offline conversion.

Pick an actively documented project

The official repository records checked on November 6, 2026 showed both projects as public and unarchived. The latest published releases observed were ebook2audiobook v26.10.1, published October 5, and Abogen v1.3.1, published February 6. These are the versions used below. Abogen also showed later repository activity; a newer branch is not the same thing as a new release. See the ebook2audiobook release and Abogen release.

Use the canonical owners, DrewThomasson/ebook2audiobook and denizsafak/abogen, when downloading. Check the release instructions for your operating system before running its installer. The ebook2audiobook documentation says its launchers may install Homebrew on macOS or Scoop on Windows to obtain missing programs. Abogen's desktop command is abogen; in this release, abogen-web and abogen-cli both start its Web UI server. The latter is not a PDF-in/audio-out batch command. These details come from the ebook2audiobook setup instructions, Abogen package entry points, and the projects' Apache and MIT license files.

What runs locally and what may call an API

Separate installation, speech generation and optional connections. Installing packages and downloading voices needs access to their distribution servers. A voice downloaded from Hugging Face can subsequently be used by a local engine; using a hosted demo instead sends the input to that hosted service.

Abogen documents a desktop option to pre-download models and voices, an option to stop Kokoro downloading from Hugging Face, and a startup update check. Its Web UI also offers LLM text normalization with a configurable OpenAI-compatible endpoint. That endpoint can be local or remote: enabling a remote endpoint changes where the normalization input is processed. Audiobookshelf integration is another optional connection to a configured server. For an offline workflow, prepare the required files and leave unneeded remote features disabled. See Abogen's interface, offline and LLM settings.

A Web UI running on your computer also needs an appropriate listening address. Abogen v1.3.1 defaults to 0.0.0.0, which binds beyond the loopback interface. For use on the same computer, this PowerShell example selects loopback before launching it:

$env:ABOGEN_HOST = '127.0.0.1'
abogen-web

This command is assembled from the documented launcher code; it has not been run for this guide. Loopback binding limits inbound access, but does not disable outgoing model downloads or API requests. Likewise, ebook2audiobook's --share option creates a public Gradio link; omit it for private local use. A practical offline check is to prepare the installation, disconnect networking and try a short non-sensitive input before committing a private document. No such network-isolation test was performed here.

Prepare the text and choose an output

First copy text from several PDF pages and compare it with the page images. Confirm that paragraphs finish before the next column begins, repeated headers are removed, and notes appear where you want them spoken. An image-only scan needs an OCR preparation step.

ebook2audiobook documents OCR support, but that does not establish accuracy on your scan. Abogen's inspected PDF extraction path calls get_text() and skips pages with no extracted text. In that path, accepting a PDF does not establish that every page was read. Inspect the text before synthesis. The relevant sources are the ebook2audiobook feature and input notes and Abogen's PDF extractor.

After installing ebook2audiobook and preparing its chosen voice, this unexecuted Windows example requests an English M4B in a separate output folder. Replace the paths with your own:

.\ebook2audiobook.cmd --headless --ebook "C:\Books\book.pdf" --language eng --output_format m4b --output_dir "C:\Books\audio"

On Linux or macOS, the documented launcher is ./ebook2audiobook.command, with the same option names. --output_format and --output_dir use underscores. Supply a single file through --ebook; do not combine it with --ebooks_dir or --text. The argument parser and format configuration document these options, PDF input, and the M4B default.

For Abogen's desktop route:

1. Run abogen, then add the PDF or your corrected text file.
2. Inspect the text and chapter/page selection. Choose a voice and speech speed.
3. Select the output format and destination. Decide whether to save chapters separately and whether subtitles are needed.
4. Start with a short selection. Open the saved audio and check its beginning and end before converting the rest.

The desktop instructions list WAV, FLAC, MP3, Opus and chaptered M4B, plus SRT or ASS subtitles. Chapter controls do not guarantee that a PDF's page breaks match its chapter boundaries. ebook2audiobook itself recommends EPUB or MOBI for automatic chapter detection. Check the resulting chapter names and sequence in your intended player; the MP3 and M4B guide explains the format choice.

Check hardware, licenses and failure recovery

Allow time and disk space for dependencies, downloaded models, temporary audio and final files. ebook2audiobook's README lists 2 GB RAM as a minimum and 8 GB as recommended, plus 1 GB VRAM minimum and 4 GB recommended for GPU use. It also lists a CPU path and warns that modern engines can be slow there. Those are the maintainer's published requirements, not a benchmark or a promise that every model fits those limits. Measure a short section on your own machine before estimating an entire book. See the hardware requirements.

Check three separate permissions: the converter's code license, the selected voice/model terms, and your right to use the document and distribute the resulting narration. The repository license alone does not answer all three. Piper's voice documentation, for example, directs readers to each voice's MODEL_CARD and warns that some voices have restrictive licenses. A locally executed workflow has hardware, electricity, storage and maintenance costs; optional hosted APIs add their own usage terms and charges.

For interruption recovery, keep the original document, corrected text, selected model, settings and completed audio together:

• ebook2audiobook: its parser documents --session for resuming an interrupted conversion. Preserve the existing session data and use its actual session identifier with the same input and settings. Check which outputs are complete before resuming; the presence of this option is not proof that every failure can be recovered without repetition.
• Abogen: the desktop documentation offers separate chapter outputs, while the Web UI exposes progress and downloadable logs. Preserve successful chapter files and use chapter selection to process the remaining work where available. These documents do not establish universal crash-safe resumption.
• After a settings or text change: verify which audio still matches the script before reusing it. A file's existence does not prove that it contains the right version or a complete ending.

These are documentation-based workflows. Installation, conversion speed, audio quality and crash recovery were not tested for this guide. If you need explicit control over saved segments and retries, continue with the Python workflow.