Still paying for Audible? audiblez turns an ebook into an audiobook in one command

1. Opening: from pain to payoff in one project

Last week a colleague handed me a 600-page English tech book. I opened the PDF and my scalp went numb: a two-hour commute, reading English on a phone in the subway, eyes watering from the strain. So I looked for an audiobook version, searched Audible, and found that either it wasn't there, or a single title cost 30-odd dollars.

I tried the traditional TTS tools myself. Either the voice sounded like a robot reading a script, or I had to spend a whole day fighting with a Python environment, only to end up with dozens of loose wav files and no chapter splits at all.

Then I found Audiblez — an open-source project with 8,400+ stars on GitHub. A 160,000-character copy of Animal Farm rendered in about 5 minutes on a Colab T4 GPU, and only about an hour on CPU. The key part: it outputs a .m4b file directly, so chapter splitting, track merging and player compatibility are all handled in one shot.

2. What it is

Audiblez is a Python tool that converts ebooks into audiobooks. Its entire pitch fits in one line: turn an .epub into a chapter-aware .m4b audiobook with a single command.

The engine underneath is Kokoro-82M, a TTS model with only 82 million parameters. Don't let the small size fool you — it was trained on less than 100 hours of audio, yet the voice is remarkably natural. It's already open source on Hugging Face.

The current release is v4, and three upgrades push it straight into "actually usable" territory:

• GUI: a wxPython desktop app, so you never have to touch the command line;
• CUDA support: enable GPU acceleration with the --cuda flag for roughly a 10x speedup;
• Language coverage: 9 languages — American English, British English, Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese and Mandarin — backed by 50+ voices.

3. Core features in depth

1. One command to generate an .m4b audiobook

The painful part of a traditional TTS pipeline isn't synthesis, it's post-processing: you split by chapter yourself, glue audio together with ffmpeg, add chapter markers, and package everything into a container your player recognizes. Audiblez wraps up the whole pipeline. Run audiblez book.epub -v af_sky and it first produces intermediate files like book_chapter_1.wav and book_chapter_2.wav, then merges them into book.m4b. That file plays directly in VLC or Apple Books, and jumping to a specific chapter works flawlessly.

2. 50+ voices × 9 languages

The naming scheme is obvious: the first letter is the language code, the second is the gender. So af_sky is the American female voice Sky, and zm_yunjian is the Chinese male voice Yunjian. The Chinese set has 8 voices, including the classic Xiaoxiao, Nini, Xiaobei and Yunjian, and its naturalness leaves concatenative TTS far behind.

3. Interactive chapter picking (--pick)

You don't always want the whole book read aloud. The --pick flag lets you interactively select chapters before conversion, skipping the foreword, table of contents and references you'd rather not hear. Saves time and effort.

4. Adjustable speed (-s)

Supports 0.5x to 2.0x. Blast through a tech book at 1.5x, savor a novel at 0.8x — one flag switches between the two.

4. Technical highlights and performance numbers

• Kokoro-82M model: just 82M parameters and under 100 hours of training data, yet it beats many larger models on naturalness scores. One of the best value open-source TTS models available;
• GPU acceleration: about 5 minutes to convert 160,000 characters on a Colab T4, roughly 600 characters/second; about 1 hour in CPU mode on an M2 MacBook Pro, roughly 60 characters/second;
• Simple architecture: pure Python, with only ffmpeg and espeak-ng as dependencies, and cross-platform support for macOS, Linux and Windows;
• Apache 2.0 + MIT dual licensing: the Kokoro model is Apache 2.0, Audiblez itself is MIT — commercial use and redistribution are both fine.

5. Where it fits and who it's for

Good fit:

• Commuters: convert your stockpile of English ebooks into audiobooks and load them onto your phone;
• Language learners: replay text in a native speaker's voice to train listening and pronunciation;
• Content creators: batch-convert blog posts, tutorials and self-published ebooks into audio to widen distribution;
• Accessibility needs: an audiobook option for blind or low-vision users, or people with reading difficulties.

Not a good fit:

• Users who need professional voice-actor-grade emotional performance (Kokoro still reads in a fairly neutral tone);
• Apple Silicon users who want native GPU acceleration (Kokoro has no MLX implementation yet, so on a Mac you're limited to CPU).

6. Comparison with alternatives

Audiblez's differentiator: it's the only option that combines automatic chapter splitting, open source, local GPU acceleration and multiple languages. edge-tts is free but has no chapters, ElevenLabs sounds top-tier but requires a subscription, and Audiblez lands right in the "good enough and free" sweet spot.

7. Getting started fast

Environment setup (macOS example):

brew install ffmpeg espeak-ng
pip install audiblez

Minimal working example:

audiblez mybook.epub -v af_sky

After a few minutes, mybook.m4b appears in the same directory. Drag it into any player and you're done.

Going further:

• GPU acceleration: audiblez mybook.epub -v af_sky --cuda — first make sure you installed the CUDA build of PyTorch;
• Faster listening: audiblez mybook.epub -v af_sky -s 1.5;
• Windows users should install inside a venv as the README suggests, to avoid dependency conflicts;
• For the GUI, just run audiblez-ui after pip install pillow wxpython.

8. Summary and next steps

Audiblez compresses "turn an ebook into an audiobook" from a tedious engineering problem into a single command. The 82M-parameter Kokoro model proves small models can produce natural speech, and with 9 languages, 50+ voices, automatic chapter splitting and CUDA acceleration, it has almost no rival among open-source TTS tools.

Recommendations: for tech books, start with af_sky or am_adam (American male) at 1.3–1.5x. For novels, bf_emma (British female) at normal speed feels better. For Chinese content, pick zf_xiaoxiao or zm_yunyang — naturalness is already close to a real person.

Links:

• GitHub: github.com/santinic/audiblez
• Model: huggingface.co/hexgrad/Kokoro-82M
• Docs: audiblez --help

One-line pitch: install Python, type one command, and your entire ebook library becomes a private Audible on the spot.