# Book to Audio

> A free PDF and ebook to audiobook converter at https://book-to-audio.skillsafe.ai/. Drop a book,
> get one audio file per chapter. The file is read in the visitor's browser and the voice (the
> open Kokoro-82M speech model) runs on the visitor's own device over WebGPU. No sign-in, no AI
> credits, no upload: the only network traffic is the one-time download of the voice model.

## Inputs

- PDF with a text layer: text in reading order (two-column pages handled), running headers and
  footers found by position and repetition, the document outline (bookmarks) for chapters. A
  scanned PDF with no text layer is detected and named; there is no OCR.
- EPUB 2 and 3: chapters from the spine documents, titles from the nav document or NCX.
- DOCX, HTML, Markdown and plain text (lines such as "Chapter 7", "PART TWO" or "IV" start chapters).
- Paste is accepted as well as files. Up to 300 MB.

## Chapters

Chosen automatically from the best evidence in the file - EPUB chapter files, then PDF bookmarks,
then heading levels, then "Chapter N" lines, else equal parts of about 3,000 words - or forced by
the visitor. Title, copyright and contents pages are found and left out by default. Every chapter
can be renamed, included or skipped, and its cleaned text shown before any audio is made.

## Cleanup (each one a switch)

Page numbers; running headers and footers; footnote marks ([12], superscripts, digits after
punctuation); web and email addresses; author-year citations; tables and code; an announced chapter
title. A "say it differently" list rewrites or skips words (one `find = replacement` per line).

## Audio

- Six English voices (American and British, female and male), speed 0.7x to 1.3x.
- About 30-second pieces are rendered in order and saved in the browser (IndexedDB) under a key of
  voice + speed + exact text, so generation resumes after a reload and a cleanup change re-renders
  only the paragraphs it altered. A live time estimate appears after the first piece.
- Per-chapter Ogg Opus (48 kbit/s, encoded by the browser's WebCodecs) or 16-bit 24 kHz WAV; all
  chapters as a ZIP with an M3U playlist; the whole book as one Ogg Opus file with chapter marks.
- Without WebGPU: chapters and cleanup still work and the browser's built-in voice reads a chapter
  aloud with the current paragraph highlighted; that voice cannot be saved as a file.

## Free example

"Try the sample book" builds a short original three-chapter PDF in the page (title, copyright and
contents pages, running header, page numbers, bookmarks) and opens it through the same reader.

## Credits

Kokoro-82M by hexgrad (Apache-2.0); pronunciation lexicon from misaki (Apache-2.0); onnxruntime-web
(MIT). The document readers and audio encoders are this app's own code.