Overlay

Turns a book into a book that reads itself: the text highlights as the audio plays. Drop in an EPUB or a PDF, get back an EPUB 3 with Media Overlays — or an M4B audiobook.

Everything runs locally on Apple Silicon. No cloud, no API keys.

What it does

• EPUB in, EPUB out. The narration is added into your book: layout, images and styles stay exactly as they were.

• PDF in. Parsed with docling, which separates body text from running heads and keeps figures and tables.

• Four voice engines, installed on demand from the app:

  engine	voices	speed	clones a voice
  Qwen3-TTS	your own sample, 9 presets	9× with batching	yes
  Kokoro	28 English	19×	no
  Kokoro RU	Sveta, Masha, Dima	18×	no
  System (macOS)	installed system voices	instant	no

  Speeds measured on an M4 Pro, relative to real time.

• Word-level highlighting when using system voices: AVSpeechSynthesizer reports the range of every spoken word, so no forced alignment is needed. Other engines highlight sentence by sentence.

Requirements

• macOS 26 or later, Apple Silicon
• ffmpeg (brew install ffmpeg)
• ~3 GB of disk per engine, downloaded on first use

Install

Download the DMG from Releases, or build it yourself:

git clone https://github.com/stepan5dol/Overlay
cd overlay
app/build.sh

Command line

python3 src/book.py BOOK.epub                    # or BOOK.pdf
python3 src/book.py BOOK.epub --voice sveta --format both
python3 src/book.py --list-voices

Nothing is required beyond the file itself: the language is detected from the book, a matching voice is picked, and missing engines are installed on demand.

Voice samples

refs/ holds reference recordings used for voice cloning, each with its transcript — cloning needs an accurate one. All shipped samples are public domain: LJSpeech (Linda Johnson) and LibriVox (Bryan Ness, Meredith Hughes).

To add your own: put name.wav (10 seconds of clean speech) and name.txt with its exact transcript into refs/.

Please only clone a voice you have permission to use.

Third-party licences

The code here is MIT. The models and libraries it downloads have their own: Qwen3-TTS and Kokoro under Apache 2.0, kokoro-ru weights under OpenRAIL, ffmpeg under GPL (used as an external binary, not linked).