How it works
Four steps, all of them on your machine. The slow one is the third.
- 01
Pick an EPUB
Drop the file on the page. It goes from your disk straight into the tab. No upload, no queue, nothing waiting on a server.
- 02
We strip out the furniture
An EPUB is a bundle of web pages, and plenty of what sits in one should never be read aloud. Page numbers, footnote markers, running heads, the contents page, the copyright notice. All of it goes, so what you hear is the book rather than the scaffolding around it.
- 03
Your machine reads it
A speech model runs in the tab, a sentence at a time, on the processor your machine already has. This is the slowest step and it is still about ten times faster than the audio it makes. Close the tab whenever you like. It keeps your place.
- 04
Download and listen
You get an M4B with chapter marks and the book's own cover art, or a folder of tagged MP3s if you prefer one file per chapter. Apple Books, Smart AudioBook Player, BookPlayer and Plex all take them.
Why it runs here instead of on a server
Speech synthesis is expensive. Every service that does this quickly is paying for hardware, which is why they need your file on their machines and an account to bill against. The moment your book is on somebody else’s server, how long it stays there is their decision rather than yours.
Running the model in the browser removes that entirely. There is no upload endpoint to send a book to and no bucket to leave it in. The cost is time, and the time is real: about an hour for a ten-hour book on an ordinary laptop.
What gets thrown away
An EPUB is a bundle of web pages, and a lot of what is in one should never be read aloud. Page numbers printed into the flow. Footnote markers, which turn into “the treaty of 1919 fourteen was signed” if you leave them in. Running heads repeated on every page. Tables, which are unnavigable as speech. The contents page and the copyright notice.
Chapter titles come from the book’s own table of contents where it has one, flattened to a single level. Sub-sections make chapter marks every forty seconds, which is noise rather than navigation.
Where the time goes
Parsing a book takes about a second. Everything after that is the speech model, running a sentence at a time. On an ordinary processor it manages roughly ten times faster than real time; on a CPU it is slower than real time, which is why we steer you towards the quicker voice for anything book-length.
Progress is written down every few seconds, so closing the tab costs you nothing. The audio is assembled at the end out of pieces that were already finished.
Nothing leaves your device
There is no upload endpoint and no storage bucket to put anything in. The voice model is WebAssembly running in your browser. Open the network tab and watch it sit idle for hours while the thing works.
The one thing that does come over the wire is the voice model, which has to arrive before it can run. It comes from our own bucket, once, and your browser keeps it. It carries nothing about you or about the book.