Why readers and authors try the open source route
It is free. The software costs nothing, and once a model is downloaded you can convert as many books as your computer can handle.
It is private. The text never leaves your machine, which matters for unpublished manuscripts and confidential documents.
It is yours to change. You can swap models, script batch jobs and feed the output into your own tools.
The voices keep improving. Open models such as Kokoro, Piper and XTTS sound far better than the system voices of a few years ago, and new ones appear often. Check each project’s README for the models it supports today.
Running it yourself vs a hosted full cast
Open source projects differ a lot from one another, so the right column describes the common pattern rather than any single tool.
| AI Audiobooks | Open source on your computer | |
|---|---|---|
| Install | Nothing; it runs in a browser | Python or Docker, model downloads, sometimes drivers |
| Hardware | Any laptop, tablet or phone | A recent CPU; a GPU speeds up larger models |
| Voices | 137 voices from Gemini Flash TTS and ElevenLabs | Whichever open models the project supports |
| Character casting | Speakers detected from dialogue tags and cast for you | Usually one voice; a few projects attempt speaker detection |
| Directions | Notes such as [whisper] plus inline tags such as [laugh] | Depends on the model; often none |
| Fixing one line | Edit and regenerate that line | Often a chapter rerun or hand editing in an audio app |
| Chapters and M4B | Built in, with named markers, tags and cover | Varies; some projects export M4B |
| Cost | Credits, about $9 to $12 per finished hour on Standard voices | Free software, plus electricity and your time |
Already have chapter MP3s from a local tool? Join them into one file in the free MP3 to M4B converter.
Setup, hardware and time per book
Most projects expect you to install Python or Docker, download one or more voice models and run commands in a terminal, and some add a simple web interface once installed. Budget an evening for the first setup, longer if drivers or dependencies fight back.
Speed comes down to the model and the machine. Lightweight models run comfortably on a CPU, while larger and more expressive ones can crawl without a GPU. Rendering a 9 hour novel locally can take anywhere from about an hour to most of a day, so time one chapter before you queue the whole book.
Character voices: the hard part to do yourself
Reading a book in one voice is the solved part. Working out who speaks each line is not: it means parsing quotation marks, reading dialogue tags such as “said Kel” and carrying the speaker through untagged back and forth. Most open source tools skip it and read everything in one voice; a few attempt it with mixed results.
To get a cast locally you can mark speakers by hand, render each part with a different voice and stitch the audio together. That is fine for a short story and a real project for a novel. A hosted multi voice audiobook tool does the attribution for you, for up to ten named characters plus an “Other voices” role.
When running it locally is the better pick
Neither route is right for everyone. Be honest about which list describes you.
Choose open source if
- You are comfortable in a terminal and enjoy setting tools up
- The book is for personal listening, or nonfiction that suits one voice
- Everything must run offline, with no text leaving your machine
- You want to convert a large personal library at no software cost
- You plan to script or customise the pipeline yourself
Choose a hosted studio if
- The book has dialogue and you want a distinct voice per character
- You need chapter markers, a cover and tags without extra tools
- You want to fix one line without rerunning a chapter
- You work on a Chromebook, a tablet or a laptop without a GPU
- Your hours are worth more than the credits for one book
Check the voice model’s licence before you sell
Open source code and open voice models are licensed separately, and some voice models allow only non-commercial use. Read the licence on the model card, not just the project’s code licence, before you publish audio made with it. Store rules on AI narration apply either way; audiobook rights explains the basics.
Compare both routes on a single chapter
- 1
Pick a test chapter
Choose one with dialogue and at least one invented name.
- 2
Render it locally
Install the project you prefer, follow its README and time the whole process, setup included.
- 3
Generate it hosted
Upload the same book to the EPUB to audiobook converter and generate that chapter with the detected cast, using the 100 gift credits (valid for 7 days, about 12 minutes on Standard; usage varies by text and voice tier) that new Google sign-ups get.
- 4
Listen, then total the hours
Compare the audio first, then the time each route took, and multiply by the chapters left.
