In depth
Why people look for an open source audiobook generator
People search for an open source audiobook generator for good reasons: it's free, the text never has to leave their computer, and they can change how it works. Look for an AI audiobook generator on GitHub and you'll find an active set of projects that turn an EPUB into an audiobook, and the open voice models behind them have improved quickly.
This page maps the projects people actually search for, what each one needs and what it produces, from their READMEs, licence files and model cards checked on October 21, 2026. Then it makes the honest case for each path: local and free, or hosted with automatic casting.
The main projects on GitHub
ebook2audiobook is the broadest. It accepts EPUB, MOBI, PDF, DOCX, TXT and more, outputs M4B with chapters and metadata by default, and switches between engines including XTTSv2, Bark, VITS and Tortoise, with optional voice cloning. It runs on CPU or GPU through a local web UI, Docker or Colab. The code is Apache-2.0, and its latest release came out on October 5, 2026.
audiblez does one thing: EPUB in, M4B out, read by Kokoro in a single voice, with ffmpeg doing the packaging. abogen takes EPUB, PDF, TXT, Markdown and subtitle files and exports MP3, FLAC, WAV or M4B with chapters, plus subtitles, and has a setting that keeps Kokoro offline. Both are MIT-licensed. epub_to_audiobook writes one MP3 per chapter for Audiobookshelf, using Azure, OpenAI or Edge voices online, or Piper and Kokoro locally.
Voices: Kokoro, Piper, XTTS and Edge TTS
Kokoro is the voice most of these tools share: audiblez, abogen and Audiobook Creator all use it, so a Kokoro TTS audiobook is the usual starting point. It's an open-weight model with 82 million parameters, released under Apache-2.0, which allows commercial use. Piper is a fast local engine: its current repository, piper1-gpl, is GPL-3.0 and says it is looking for maintainers, and every Piper voice ships with its own model card and licence, some of them restrictive.
Coqui's XTTS-v2 can clone a voice from a six-second clip in 17 languages, but its Coqui Public Model License allows only non-commercial use of the model and its output. And an edge tts audiobook isn't local at all: the edge-tts README says it uses Microsoft Edge's online text-to-speech service, even though the code itself is open source.
Character voices: who attempts speaker attribution
Most projects read the whole book in one voice. ebook2audiobook can switch voices, but you mark the passages yourself with voice tags. Audiobook Creator is the one that tries to do it automatically: a language model you run, through any OpenAI-compatible API, first lists the characters with their age and gender, then attributes each line, and the tool assigns voices by gender. Its README suggests a context window of at least 20,000 tokens.
That's the problem FullCastBook solves on its servers. Take a line like OFFICER REYES: [firm, tired] Step away from the car. A one-voice tool reads it like the narration around it; a full cast gives Reyes her own voice and keeps it for the rest of the book.
Setup, hardware and time per book
Expect Python or Docker, ffmpeg for M4B (calibre for Audiobook Creator) and some command-line work. A GPU is optional for most projects but changes the wait. audiblez's README reports about 600 characters a second on a Colab T4 GPU and about 60 a second on an M2 MacBook Pro's CPU. By our arithmetic, an 80,000-word novel is roughly 470,000 characters, so that's around 13 minutes on that GPU or two to three hours on that CPU, before any re-takes.
Hardware support varies by project. Audiobook Creator says Apple Silicon GPUs aren't supported, ebook2audiobook lists Apple's MPS alongside CUDA and ROCm, and abogen supports AMD GPUs on Linux only. If you'd rather try a free hosted model by hand, see the Google AI Studio audiobook walkthrough.
Where a hosted full cast saves work
FullCastBook trades the free and offline parts for less work. There's nothing to install: upload an EPUB, a PDF with a text layer, a DOCX, TXT or Markdown file, and chapters and speakers are detected for you. Each character gets a distinct voice next to the narrator, you can direct a line with [whisper] or [sigh], regenerate one line instead of a chapter, and export an M4B with chapter markers, one MP3 or a ZIP of chapters.
Open source still wins on cost, privacy and control. It can run offline, it's free beyond your hardware, you can read and change the code, and ebook2audiobook can clone a voice, which we don't offer. Your text also stays on your machine; with FullCastBook, it's sent to our voice providers, Google's Gemini TTS for standard voices and ElevenLabs for premium ones.
Which path to choose
Choose a local AI audiobook generator if you're comfortable with Python or Docker, want one good voice for free, or need everything offline, and check the licence of the voice before you sell anything. Choose a hosted full cast if the book is dialogue-heavy, you'd rather not maintain a toolchain, or you want chapters and character voices handled in one place, in the audiobook studio.
For the wider field, including voice studios and store programs, see the best TTS for audiobooks guide; for the format itself, the M4B audiobook maker.
