A voice you'd want to listen to.
Listen to any web page with natural AI voices, entirely on your device.
Install, click, listen. English only for now (American and British voices). No account, no server, and nothing you read ever leaves your browser. Lector runs the Kokoro 82M speech model locally with WebGPU (or WASM on machines without a GPU).
Inspiration: Voicebox by Jamie Pine. Lector can later connect to a local Voicebox as an alternative voice engine.
- Natural voices. 28 American and British English voices with instant previews and favourites.
- Reads like a person. Speaks in whole sentences, even in all-lowercase posts, and says numbers the way you would: "$2.5M" is "two point five million dollars", "2026-10-07" is "October 7, 2026", "World War II" is "World War Two".
- Follow along, word by word. The sentence being read is highlighted on the page and the spoken words are underlined as they're read, so you never lose your place. Pause and keep reading yourself any time. Turn it off with Show on page.
- Start anywhere. Select text → right-click → Listen from here, press
Alt+Shift+H, or Alt+click a paragraph while listening. - Full control. Seek bar with time left, previous/next paragraph, 0.75×–2× speed without pitch change, media keys and your computer's Now Playing controls.
- Light on your computer. Uses your graphics chip when it can, multi-threaded CPU otherwise, and frees memory when you stop. Choose Voice quality in Settings if you want a smaller download.
- Accessible. Works fully with a keyboard and screen readers, supports high-contrast mode and reduced motion, follows your light or dark theme.
- Private by design. Uses
activeTab, so it only touches a page when you ask it to. The only network access is the one-time model download from Hugging Face. See PRIVACY.md.
Chrome Web Store: Lector has been submitted and is waiting for Google's review. The store link will appear here once it's approved.
Until then, install the release build:
- Download lector.zip from the latest release.
- Unzip it.
- Open
chrome://extensions, turn on Developer mode (top right), click Load unpacked, and choose the unzipped folder.
Chrome may remind you that a developer-mode extension is installed; that's expected for builds from outside the store. When the store version is live, remove this one and install Lector from the store to get automatic updates.
The first time you listen, the voice model downloads once (about 90–330 MB depending on your computer) and is cached by the browser. After that, Lector works offline.
From source:
git clone https://github.com/harshmathurx/lector.git
cd lector
bun install
bun run build # then Load unpacked → dist/| Do this | To get this |
|---|---|
| Click the toolbar icon → Listen to this page | Read the whole article |
Alt+Shift+R |
Start, or play/pause, from any tab |
Alt+Shift+H |
Listen from your selection, cursor or the focused paragraph |
Alt+Shift+. / Alt+Shift+, |
Next / previous paragraph |
| Select text → right-click → Listen from here | Start from that spot |
| Select text → right-click → Listen to selection | Read just the selection |
Alt+click a paragraph while listening |
Jump there |
| Media keys / Now Playing | Play/pause, previous/next paragraph, skip 15s |
On a Mac, Alt is the Option key. Shortcuts can be changed at chrome://extensions/shortcuts.
Language support: English only. Lector reads English pages with 28 American and British voices. Other languages are not supported yet and will not read well.
Popup ──────────────┐ polls state, sends commands
Content script ─────┤ extracts the article, highlights the spoken sentence
Background worker ──┤ lifecycle, routing, session recovery
Offscreen document ─┘ Kokoro TTS + Web Audio playback
Articles are split into whole-sentence segments (src/shared/chunker.ts); over-long sentences break at the most natural phrase boundary, and the engine splits anything still over Kokoro's 510-token limit instead of truncating it. Each segment is then normalized for speech (src/shared/normalize.ts: currency, dates, units, Roman numerals, acronyms). A small lookahead pump generates a few segments ahead of the playhead; each segment's pause is baked in as trailing silence, so pause, seek and skip are all segment-based. Generated audio is cached in IndexedDB. The heavy ONNX/Kokoro code is loaded lazily so a failure surfaces as an error message instead of a silently dead page. See CLAUDE.md for the full architecture and design decisions.
bun install
bun run build # build to dist/
bun run watch # rebuild on change
bun run typecheck # tsc --noEmit
bun test # unit tests
bun run package # build the store/release zip into release/scripts/e2e.mjs is a real-browser end-to-end test (loads the built extension in Chromium with the real model) and scripts/bench.mjs measures start time, speed and memory; their headers explain the setup. See CONTRIBUTING.md.
Lector is MIT licensed (LICENSE). It builds on Kokoro (Apache-2.0), kokoro-js and Transformers.js (Apache-2.0), ONNX Runtime Web (MIT) and Mozilla Readability (Apache-2.0). Type is set in Gloock and Instrument Sans (SIL OFL 1.1, bundled in static/fonts/). See THIRD_PARTY_NOTICES.md.