Media
4 items
Description
Read selected text aloud with a natural AI voice. Private on-device text to speech (TTS) — works offline, fast on Apple Silicon. Select any text on any page and hear it read aloud in a natural neural voice — with the sentence being spoken highlighted so you never lose your place. SelectSpeak is different from other read-aloud extensions in one important way: the AI voice runs entirely on your computer. Your text is never sent to a server. There is no account, no subscription, no tracking, and no word limit — and our build pipeline enforces it: after the one-time voice download, the extension makes zero network requests. HOW IT WORKS Select text — a small Listen button appears (or press Cmd+Shift+Period / Ctrl+Shift+Period). A neural voice (Kokoro AI) starts reading, and each sentence is highlighted on the page as it's spoken, with auto-scroll. A mini player gives you play/pause, skip ±10s, a progress bar, voice picker, and a 0.5×–4× speed slider. WHY PEOPLE USE IT • Dyslexia and reading fatigue — listen while you follow the highlight • Proofreading your own writing by ear • Getting through long articles, docs, and papers faster • Listening at high speed: unlike ordinary TTS, speech stays clear and articulate even at 3×–4×, thanks to pitch-preserving time-stretching that protects the natural pauses between words WORKS WHERE YOU READ • Any regular web page, with live sentence highlighting • Gmail, comment boxes, and rich-text editors • Text fields and text areas (never password fields) • Google Docs (best effort — no highlighting) VOICES & SPEED • 11 natural Kokoro voices — US and UK, female and male • 0.5× to 4× speed, changeable mid-playback without losing your spot PRIVATE BY DESIGN • 100% on-device text to speech — your text never leaves your machine • No account, no sign-up, no analytics, no telemetry • Works fully offline after the one-time voice download • Free and open source FAST ON MODERN HARDWARE SelectSpeak uses your GPU through WebGPU when available — on Apple Silicon Macs (M1/M2/M3/M4) generation is near-instant. On machines without WebGPU it falls back to a smaller CPU model that runs near real time. GOOD TO KNOW First use downloads the voice model (~90 MB, or ~330 MB on WebGPU). This happens once; the model is cached locally and everything runs offline afterwards. • Requires Chrome 120+. PERMISSIONS, EXPLAINED "Read data on websites you visit": needed to read the text you select and draw the highlight. Nothing is collected or transmitted. • huggingface.co: the only network access, used once to download the voice model, verified against checksums pinned at build time.