Read selected text — or a whole article — aloud with Kokoro TTS, entirely on your machine.
A Chromium browser extension (Chrome / Brave / Edge) that narrates web pages with a neural TTS voice. Select text and press a shortcut, or read the whole article — a floating control bar gives you play/pause, a seek scrubber, speed, and voice. Audio streams: playback starts on the first sentence while the rest synthesizes.
Everything is local. Nothing leaves your machine — inference runs in your browser (WebGPU/WASM) or against a local Docker server you control.
- Read a selection — select text →
Cmd/Ctrl+Shift+S(or right-click → Read selection aloud). - Read the whole article —
Option/Ctrl+Shift+A(or right-click → Read this article). Extracts the main text (drops nav/ads/comments) with Mozilla Readability; poorly-extracting pages fall back to a manual-selection hint instead of reading garbage. - Floating control bar — play/pause, seek, time, speed (0.75–2×), voice, a live engine marker (🟢 WebGPU / 🟡 WASM / 🔵 Server), and close.
- Streaming playback — starts on the first sentence; each sentence is MP3-encoded on the fly into one pipeline shared by every engine.
- Three engines, one boundary — pick in settings; speed and voice persist between reads.
| Engine | Marker | Needs | Notes |
|---|---|---|---|
| Browser · WebGPU | 🟢 | A GPU adapter | Fastest. Downloads the model once (~300 MB), then browser-cached. |
| Docker server | 🔵 | A local Kokoro server | OpenAI-compatible local server. See server/. |
| Browser · WASM | 🟡 | Nothing | Always available, but below real-time — great for short selections, buffers on long reads. |
Automatic (default) uses the fastest available: WebGPU → Docker server → WASM. A dead engine is never auto-selected, and an explicitly-chosen engine that's unavailable surfaces an error rather than silently falling back.
Distribution is load-unpacked / open-source — not an extension-store build.
git clone https://github.com/HasanResul/audio-reader.git
cd audio-reader/extension
npm install
npm run build # required only for the in-browser WebGPU/WASM enginesThen load it:
- Open
chrome://extensions(orbrave://extensions/edge://extensions). - Toggle Developer mode.
- Load unpacked → select the
extension/folder. - Toolbar icon → Settings: pick an Engine and a voice/default speed.
- (Optional) Confirm the shortcuts at
chrome://extensions/shortcuts.
The build step bundles the in-browser engine (kokoro-js + onnxruntime-web + an MP3 encoder) and vendors the ORT wasm — MV3 forbids remote script. If you only ever use the Docker server engine you can skip the build; the extension loads without the bundle and just won't offer the browser engines.
The optional Docker server engine is documented in server/README.md.
One extension, one engine boundary: everything except "produce audio from text" is
shared across engines. A hidden offscreen document runs WebGPU + audio playback and is
the only context allowed to reach localhost; a shared, engine-agnostic player streams
MP3 chunks from whichever engine produced them.
Full architecture, file-by-file breakdown, and known limitations live in
extension/README.md. The design history, research, and
benchmarks are under docs/.
- Chromium-based browser (Chrome / Brave / Edge) — the in-browser engine needs WebGPU/WASM as shipped by Chromium.
- Node.js + npm (for the in-browser engine build).
- Docker (optional, only for the Docker server engine).
Issues and PRs welcome — see CONTRIBUTING.md. For security reports, see SECURITY.md.
MIT. This project vendors and bundles third-party components under their own licenses — see THIRD-PARTY-NOTICES.md.