Sawt

صوت (sawt): Arabic for "voice" or "sound".

Listen to any web page, or turn a whole book into an audiobook, with a natural voice that runs on your own computer.

Free, private, offline. No account, no subscription, no usage limits, and nothing you read is ever sent to a server: the voice model (Kokoro, 82M parameters) runs locally on your CPU.

• Read any article aloud from the browser: one click or Alt+R. Menus, ads and sidebars are skipped, the sentence being spoken is highlighted, and every page resumes where you left off.
• Meanings of hard words as you listen. When a sentence holds a word like "myopic", a small card beside the article explains it, from an offline dictionary or from the article itself ("maximum allowable operating pressure (MAOP)"). Reading never pauses.
• Follow along word by word, optionally: the word being spoken can be gently magnified.
• Convert books (EPUB, PDF, TXT) into .m4b audiobooks with chapter markers, mastered to audiobook loudness standards.
• 28 English voices, American and British, with adjustable speed.

Status: v0.2, early. It works well for the author daily; expect rough edges. Bug reports and contributions are very welcome. See CONTRIBUTING.md.

How it works

 Browser extension                        Your computer
┌───────────────────────┐   article text   ┌──────────────────────────────┐
│ finds the article     │ ───────────────▶ │ sawt serve              │
│ plays + highlights    │ ◀─────────────── │ Kokoro voice model (CPU)     │
└───────────────────────┘   sentence audio  └──────────────────────────────┘
                     only over 127.0.0.1 - never the internet

The extension can't run a voice model by itself, so it talks to a small engine you run locally. The engine keeps the model loaded, synthesizes a few sentences ahead of playback, and caches the audio, so re-opening an article plays instantly.

Install

1. Install the engine. You need uv, the Python installer (skip this if uv --version already works):

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh
# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Then:

uv tool install sawt
sawt agent install   # downloads the voice model once (~350MB), then runs in the background

From then on the engine starts by itself when you log in and restarts if it stops, so there is no window to keep open. sawt agent status checks on it and sawt agent uninstall removes it. To run it only when you want, use sawt serve and keep that terminal open instead.

After uv tool upgrade sawt, run sawt agent install again to restart on the new version.

2. Add the browser extension from the Chrome Web Store. Its setup page checks that the engine is running.

3. Open any article and press Alt+R.

Usage

Reading web pages

More: docs/extension.md (highlighting, resume, limits).

Converting books

Needs FFmpeg on your PATH.

sawt extract ~/Downloads/mybook.epub     # -> chapters
sawt inspect                             # check what extraction produced
sawt synth                               # synthesize (hours for a novel; resumable)
sawt build                               # -> ~/Sawt/out/<book>.m4b

Or several at once, unattended: sawt batch books/*.epub --voice af_bella --speed 0.85. The full guide (voices, pronunciation fixes, trimming front matter, quality checks): docs/books.md.

Run sawt with no arguments for every command.

Configuration

Platform support

English only for now. Firefox and Safari are not supported yet.

Privacy and security

• The extension reads a page only when you ask, and sends its text only to 127.0.0.1.
• The engine makes network requests only to download two files once, on first start: the voice model and the word glossary, both from GitHub and both verified against pinned SHA-256 hashes.
• The engine refuses requests from web pages and from other extensions. Details in SECURITY.md; the full policy is in PRIVACY.md.

Troubleshooting

Roadmap

• A desktop app that bundles the engine, so no terminal is needed
• Firefox support
• More languages (Kokoro has voices for Spanish, French, Hindi, Italian, Japanese, Portuguese and Chinese)

Ideas and votes welcome in issues.

Contributing

Pull requests are welcome, from typo fixes to new features. Start with CONTRIBUTING.md (setup takes a couple of minutes) and docs/architecture.md for a map of the code. Issues marked good first issue are a gentle way in.

Credits

• Kokoro by hexgrad, the voice model (Apache-2.0)
• kokoro-onnx, which runs it on the CPU (MIT)
• Readability by Mozilla, which finds the article on a page (Apache-2.0)
• Word meanings come from Wiktionary (CC BY-SA 4.0), via the Kaikki.org extraction; word frequencies from wordfreq
• Pronunciation falls back to espeak-ng (GPL-3.0), installed as a dependency of kokoro-onnx

Convert books you own, public-domain texts and your own writing. Distributing a synthesized copy of an in-copyright book is not allowed.

Licence

MIT