audio-book-generator PDFs to EPUBs to narrated audiobooks

audio-book-generator

Two tools in one, and you choose what you want:

  1. PDF → EPUB: turn a PDF, text-based or a scanned book, into a proper EPUB with chapters, a table of contents and a cover. Scanned pages are read with OCR and cleaned up: running headers, page numbers and OCR noise removed, split words rejoined, paragraphs rebuilt.
  2. EPUB (or PDF) → audiobook: a narration that sounds read, not recited, as a single .m4b with chapters and cover art, or one MP3 per chapter.

Everything runs on your own computer with free voices: no account, no internet once the voices are downloaded.

tests Python OCR Voices Platform License

Features

PDF to EPUB

Audiobooks

How it makes synthetic speech sound human

A speech model reads exactly what it’s given. Most of the “robot” in robot narration comes from what happens around the model, so this project works on that:

Problem What a naive converter does What this does
Mr. Holmes “Mister.” (full stop: a sentence-ending pause mid-sentence) “Mister Holmes”
in 1881 “one thousand eight hundred eighty-one” “eighteen eighty-one”
CHAPTER IV. “C-H-A-P-T-E-R I-V” or “Chapter I V” “Chapter Four.”
£5 10s 6d, 4½ per cent, the 1890s digits and symbols “five pounds, ten shillings and six pence”, “four and a half per cent”, “the eighteen nineties”
Henry VIII, World War II “Henry V I I I” “Henry the eighth”, “World War two”
Pauses the same gap everywhere, like a metronome 0.30–0.46 s between sentences, 0.62–0.92 s between paragraphs, 0.42–0.62 s in quick dialogue, about 2 s at a scene break, varied within each range
Pace one fixed speed slightly slower than the model’s default (0.94), headings slower still, small drift between paragraphs, dialogue a shade quicker
Phrasing one sentence at a time (the model sounds weak on short phrases) or whole paragraphs (it rushes long input) sentences grouped into pieces of 100–330 characters, the length the model reads best; long sentences split at a clause
Silence dead digital silence between sentences faint room tone (-72 dBFS) underneath, like a quiet recording booth
Joins clicks where pieces meet 8 ms fades at every join

All variation is seeded from the text itself, so running the same book twice gives the same audio.

Voices

The grades are the model author’s estimate of the quality and amount of training audio behind each voice. Voices sound different to different people, so listen to a few with abgen preview.

Voice   Accent Grade
af_heart (female, default) female American A
af_bella female American A-
af_nicole female American B-
bf_emma (female-uk) female British B-
am_michael (male) male American C+
am_fenrir, am_puck male American C+
bm_george (male-uk), bm_fable male British C

abgen voices lists all 28. The male voices are graded lower than the best female ones; a blend such as am_michael:60,am_fenrir:40 often sounds smoother than either alone.

Requirements

No GPU is needed.

Installation

  1. Install the system tools:

    sudo apt install ffmpeg poppler-utils tesseract-ocr                              # Debian, Ubuntu, Mint
sudo dnf install ffmpeg-free poppler-utils tesseract tesseract-langpack-eng      # Fedora
sudo pacman -S ffmpeg poppler tesseract tesseract-data-eng                       # Arch

    Add Tesseract language packs for books in other languages (e.g. tesseract-ocr-fra / tesseract-langpack-fra).

  2. Get the code and run the installer:

    git clone https://github.com/y-v-j/audio-book-generator.git
cd audio-book-generator
./install.sh

This creates the conda environment py-ebook (Python 3.12, conda-forge only), installs the project into it, adds an abgen command to ~/.local/bin, and downloads the model files (verified by SHA-256) to ~/.local/share/audio-book-generator/models. Options: --env NAME, --no-models, --fast-model (also fetch the smaller int8 model).

Check it works with abgen voices and abgen preview --voice male.

Manual installation, with conda:

conda create -n py-ebook --override-channels -c conda-forge python=3.12 pip
conda activate py-ebook
pip install -e '.[test]'
abgen models

or with a plain virtual environment (Python 3.10–3.13):

python3.12 -m venv .venv
. .venv/bin/activate
pip install -e '.[test]'
abgen models

Where files go

What Where
Voice model (about 350 MB) ~/.local/share/audio-book-generator/models (or $ABGEN_MODEL_DIR)
Work in progress and OCR results ~/.cache/audio-book-generator (finished books clean up after themselves; --keep-cache keeps them)
Your pronunciation fixes ~/.config/audio-book-generator/lexicon.txt
EPUBs and audiobooks next to the input file, or wherever -o says

Uninstall

conda env remove -n py-ebook
rm ~/.local/bin/abgen
rm -r ~/.local/share/audio-book-generator ~/.cache/audio-book-generator ~/.config/audio-book-generator

Usage

PDF to EPUB

abgen info scan.pdf                            # pages read, chapters found, estimated audio length
abgen epub scan.pdf                            # -> scan.epub next to the PDF
abgen epub scan.pdf -o "My Book.epub" --title "My Book" --author "A. Writer"
abgen epub scan.pdf --ocr always --ocr-lang eng+fra --dpi 300

The first command on a scanned PDF runs the OCR (about 1.5 s per page on a 4-core laptop) and caches it; later commands on the same file are instant.

PDF option Meaning
--ocr auto\|always\|never auto (default): OCR only pages without text; always: OCR every page (for PDFs whose text layer is poor); never: text only
--ocr-lang Tesseract language(s), e.g. eng, fra, eng+deu (default eng)
--dpi resolution for OCR (default 300; 200 is faster on clean scans)
--ocr-workers pages read in parallel (default: number of CPU threads)
--title, --author for PDFs that don’t say what they are (also work for EPUBs)

These options work with info, epub, text and convert.

Audiobooks

abgen voices                                   # list voices
abgen preview --voice male                     # hear a voice (saves and plays a short WAV)
abgen preview --voice bf_emma --speed 0.9 --text "It was the best of times."
abgen info book.epub                           # chapters, word counts, estimated length
abgen text book.epub -c 3                      # exactly what will be spoken, with every pause
abgen convert book.epub                        # -> book.m4b next to the EPUB
abgen convert book.epub --voice male-uk -o ~/Audiobooks/book.m4b
abgen convert book.epub -c 1-3                 # only chapters 1 to 3 (numbers from 'abgen info')
abgen convert book.epub -f mp3                 # one MP3 per chapter: "book (audiobook)/001 - Title.mp3"
abgen convert scan.pdf --save-epub             # straight from a PDF, keeping the EPUB too

Options for preview, text and convert:

Option Meaning
-v, --voice voice name, alias, or blend a:60,b:40 (default female)
-s, --speed reading speed, 0.5–2.0; 1.0 is the model’s own pace (default 0.94)
--plain no variation in pauses and pace
--fast the smaller int8 model: about 15% faster, slightly lower quality
--lexicon FILE pronunciation fixes (default ~/.config/audio-book-generator/lexicon.txt if it exists)
--no-intro don’t announce the title and author at the start and “End of …” at the end
--room-tone DB background level in dBFS, or off (default -72)

Only for convert: -c/--chapters (e.g. 1-5,8,10-), -f/--format m4b|mp3, -o/--output, -w/--workers N, --bitrate (default 64k AAC, plenty for speech), --keep-cache, --save-epub [FILE] (when the input is a PDF).

How long it takes

Synthesis runs on the CPU. On a 4-core laptop processor with four workers it takes about 0.6 seconds of computing per second of speech, so a 10-hour book takes roughly 6 hours. Long books are best left running overnight. If you stop the conversion, run the same command again and it continues from where it stopped.

Pronunciation fixes

Copy lexicon.example.txt to ~/.config/audio-book-generator/lexicon.txt and add lines like:

Hermione = Her-my-oh-nee
221B = two twenty-one B

Spell the word the way it sounds. Check it with abgen preview --text "Hermione smiled.".

Troubleshooting

Project layout

abgen/
  cli.py        command line (voices, preview, info, epub, text, convert, models)
  epub.py       reads the EPUB: table of contents, chapters, paragraphs, headings, scene breaks
  pdf.py        reads a PDF: text layer or OCR, clean-up, paragraphs, headings, chapters
  epubwrite.py  writes a valid EPUB 3 (with EPUB 2 table of contents)
  textnorm.py   rewrites text for speech (abbreviations, numbers, money, Roman numerals, lexicon)
  segment.py    sentences and model-sized pieces
  prosody.py    pauses and pace
  voices.py     voice list, grades, aliases and blends
  engine.py     Kokoro synthesis in parallel worker processes
  pipeline.py   plan -> synthesize (resumable) -> assemble chapters -> encode
  audio.py      fades, room tone, loudness, M4B/MP3 encoding with ffmpeg
  models.py     model download and SHA-256 verification
tests/          unit tests, plus slow tests that run OCR and the speech model (pytest -m slow)
install.sh      environment, command and model set-up

Run the tests with pytest -m "not slow" (fast, no model needed) or pytest (also runs OCR and a real two-chapter conversion; needs the model files, ffmpeg, tesseract and ImageMagick).

Use responsibly

Convert books you own or that are free to use, such as public-domain titles from Project Gutenberg or Standard Ebooks. EPUBs and audiobooks you make from copyrighted books are for your own use; sharing them may break copyright law where you live. The tool doesn’t remove DRM and won’t open DRM-protected EPUBs.

Acknowledgements

This project stands on the work of others:

None of these are included in this repository. Python packages are installed by pip and the voice model is downloaded on first use; each keeps its own licence. espeak-ng and phonemizer are GPL-licensed, so a bundle that ships them together with this code (for example a single executable) must follow the GPL.

License

MIT © 2026 Yogesh