DEV Community

Cover image for How to Convert an EPUB to an M4B Audiobook in a Cloned Voice (No GPU, Open Source)
Miragea
Miragea

Posted on

How to Convert an EPUB to an M4B Audiobook in a Cloned Voice (No GPU, Open Source)

miragea-audiobook banner

If you have searched for "EPUB to audiobook" you have probably met tools that want PyTorch, a few gigabytes of model weights and a GPU. This tutorial does the same job from a plain command line: miragea-audiobook reads an EPUB, Markdown or text file, narrates it through the Miragea API in a voice cloned from a short sample, and gives you one chaptered .m4b.

What you need

  • Python 3.10+ and ffmpeg (brew install ffmpeg or apt install ffmpeg)
  • A Miragea API key from https://miragea.ai/workspace/api
  • A clean 5-30 second recording of the voice you want, hosted at a public https URL, and the exact words spoken in it

1. Install and set the key

pipx install miragea-audiobook
export MIRAGEA_API_KEY=...
Enter fullscreen mode Exit fullscreen mode

2. Check what it found (offline)

miragea-audiobook chapters book.epub
miragea-audiobook plan book.epub
Enter fullscreen mode Exit fullscreen mode

EPUB chapters come from the spine and each file's first heading. Covers and copyright stubs under 20 words are skipped. For Markdown, each ## heading is a chapter; for plain text, lines starting with "Chapter", "Part", "Prologue" and similar.

How the pipeline works: read, chunk, voice, assemble, package

3. Make the audiobook

miragea-audiobook make book.epub \
  --voice 'https://example.com/my-voice.wav|Exact words spoken in that sample.' --split
Enter fullscreen mode Exit fullscreen mode

The text is cut into whole-sentence chunks (900 characters by default), voiced in parallel, and cached by request hash. If something fails halfway, run the same command again: finished chunks are reused and not billed twice.

Terminal capture of a real run producing a two-chapter M4B

4. What you get

the-salt-road.m4b with chapter markers, a .cue sheet, a .chapters.txt list ready for YouTube descriptions, a JSON manifest and one MP3 per chapter. Here is the real waveform of the demo book, with the markers at the chapter starts:

Waveform of the demo audiobook with chapter markers

Listen to the real demo: https://codeberg.org/miragea-voice/miragea-audiobook/raw/branch/main/examples/demo/the-salt-road.m4b

I ran chapter 1 back through speech recognition and it matched the source text word for word.

Tips

  • Use --preview first on a big book: it builds the whole file offline with placeholder tones so you can check chapters and timing for free.
  • --chapters 1-3 voices just a few chapters while you audition a voice.
  • Audiobookshelf, Apple Books and VLC show the chapter list automatically.
  • Only clone a voice you have the right to use.

Source (MIT): https://codeberg.org/miragea-voice/miragea-audiobook. Package: https://pypi.org/project/miragea-audiobook/

Top comments (0)