Speak freely. Type instantly. 100% private & offline.
Free, open-source voice dictation for Chrome. Whisper runs entirely inside your browser through Transformers.js, ONNX Runtime Web and WebGPU. Your microphone audio never leaves your machine — there is no server to send it to.
Install · How it works · Build from source · Contributing · Privacy
Put your caret in any text field on any site, press Ctrl+Shift+D, and talk. The words
appear where the caret is — in Gmail, Notion, Slack, GitHub issues, a Jira ticket, a
search box, a <textarea> on a form nobody has updated since 2011.
Everything happens locally. The only network request LocalDictate ever makes is the one-time model download from Hugging Face. After that you can turn off Wi-Fi and keep dictating.
Left: the popup transport panel. Right: the in-page listening bar.
| Whisper, on device | Tiny, Base and Small, converted to ONNX and quantised for the browser. |
| WebGPU with a real fallback | Detects an adapter and uses it; drops to WebAssembly automatically, and tells you which one is running. |
| Near real-time | Utterances are segmented by voice activity detection and decoded as you pause, with streaming partials in between. |
| Types where you're typing | Native value setters for inputs, insertText for contenteditable, clipboard fallback for canvas editors. |
| Silence never reaches the model | An energy gate with an adaptive noise floor throws away room tone before it costs you any compute. |
| Clean output | Fillers removed, sentences capitalised, punctuation spaced. Spoken commands — "new paragraph", "comma", "question mark" — work out of the box. |
| Your vocabulary | Word replacements for names and acronyms Whisper mangles. |
| Cached and offline | Weights live in IndexedDB. Second launch is instant and needs no connection. |
| Yours to inspect | MIT licensed, no build obfuscation, no analytics, no accounts. |
Coming soon — the listing is in review. Until then, load it unpacked.
git clone https://github.com/wushu75/LocalDictate.git
cd LocalDictate
npm install # fetches Transformers.js + ONNX Runtime Web
npm run build # copies them into lib/- Open
chrome://extensions. - Turn on Developer mode (top right).
- Click Load unpacked and select the
LocalDictatefolder. - A welcome tab opens. Grant microphone access, then download a model.
The lib/ folder is intentionally not committed — it holds upstream build artefacts.
npm run build is what puts them there, and the extension will not start without it.
- Grant the microphone. Chrome asks once, for the extension itself.
- Pick a model. Base is the right first choice. Tiny if you are on an old laptop, Small if you have a discrete GPU and care about accuracy.
- Wait for the download. 78 MB to 490 MB depending on the model, once, ever.
- Click into a text field and press
Ctrl+Shift+D.
Prefer holding a key? Settings → Dictation → Trigger → Push to talk. The default hold key is Right Ctrl, and you can rebind it to anything.
┌───────────────┐ hotkey ┌──────────────────┐
│ content script│────────────▶│ service worker │ routes, never touches audio
│ focus + HUD │◀────────────│ │
└───────────────┘ text └────────┬─────────┘
▲ │ chrome.offscreen
│ insertText ▼
│ ┌──────────────────────────────┐
│ │ offscreen document │
│ │ getUserMedia → AudioWorklet │
│ │ 16 kHz mono → VAD → segments│
│ └───────────┬──────────────────┘
│ │ Float32Array (transferred)
│ ▼
│ ┌──────────────────────────────┐
└─────────────────────│ module worker │
│ Transformers.js + ORT Web │
│ Whisper on WebGPU or WASM │
└──────────────────────────────┘
Why an offscreen document? MV3 service workers can't hold a MediaStream or an
AudioContext, and they get torn down mid-sentence. The offscreen document owns the
capture graph and the worker; the service worker only routes messages.
Why a separate worker for inference? Decoding a 20-second utterance on WASM can block a thread for seconds. Keeping it off the document's main thread keeps the level meter and the HUD responsive while Whisper works.
Segmentation. Audio arrives as 1024-sample frames (64 ms). Each frame gets an RMS measurement against an adaptive noise floor. Speech opens an utterance; 700 ms of silence closes it and sends it for a final decode. While an utterance is open, a partial decode runs every ~1.4 s so you can see the words forming.
More detail in docs/ARCHITECTURE.md.
Rough figures for one 8-second utterance, measured end to end on a 2023 laptop. Yours will differ; the ratios are what matter.
| Model | WebGPU | WASM (CPU) | Download |
|---|---|---|---|
| Tiny | ~0.4 s | ~1.6 s | ~78 MB |
| Base | ~0.7 s | ~3.4 s | ~148 MB |
| Small | ~1.8 s | ~11 s | ~490 MB |
The first utterance after loading is slower — that's shader and graph compilation. It happens once per session, and LocalDictate runs a silent warm-up pass to absorb most of it.
npm install # dependencies
npm run vendor # copy transformers.js + ort wasm into lib/
npm run icons # regenerate icons/*.png from scripts/make-icons.py (needs Pillow)
npm run package # produce dist/localdictate-<version>.zip for the Web Store
npm run clean # remove dist/ and vendored lib filesThere is no bundler and no transpiler. Every file you read in src/ is the file the
browser runs, which is the point: a privacy tool should be auditable without a build
graph in the way.
LocalDictate/
├── manifest.json Manifest V3
├── icons/ 16 / 32 / 48 / 128 / 512 px + SVG source
├── lib/ vendored (git-ignored, created by npm run build)
│ ├── transformers.js
│ └── ort/ onnxruntime-web wasm binaries
├── src/
│ ├── background/service-worker.js hotkeys, offscreen lifecycle, routing
│ ├── offscreen/ mic capture, VAD, worker orchestration
│ │ ├── offscreen.html
│ │ ├── offscreen.js
│ │ └── pcm-processor.js AudioWorklet
│ ├── worker/whisper-worker.js Transformers.js pipeline
│ ├── content/ focus tracking, HUD, text insertion
│ ├── popup/ transport panel, model picker, transcript
│ ├── options/ full settings
│ ├── permission/ microphone grant + onboarding
│ └── common/ constants, storage, VAD, post-processing, cache
├── scripts/ vendor, package, icon generation
├── store/ Chrome Web Store listing copy and assets
└── docs/ architecture notes and screenshots
The short version: no audio, no transcripts and no usage data ever leave your device. There is no backend, no account, no analytics SDK and no remote logging in this codebase.
The one network request is the model download from huggingface.co, declared in
host_permissions. It happens once per model and is cached in IndexedDB.
Full statement: PRIVACY.md. If you find anything in this repository that contradicts it, that is a security bug — please open an issue immediately.
Nothing appears in the page. Click into the field first. LocalDictate types into the element that had focus when you started. If the site uses a canvas editor (Google Docs), the text is copied to your clipboard instead and a toast tells you to paste.
Google Docs. Docs renders text on a canvas with no real editable DOM. No extension can type into it directly. Use the clipboard fallback, or dictate in the comment box, which is a normal contenteditable.
"WebGPU is unavailable." You're on WASM. It works, it's just slower — stick to Tiny or
Base. Chrome exposes WebGPU on most modern hardware; check chrome://gpu.
The first sentence is missing words. Voice activity detection needs ~160 ms of speech to open an utterance. Lower the sensitivity slider, or use push-to-talk, which ignores the gate while the key is held.
It's transcribing my music / air conditioner. Raise the sensitivity slider. Whisper will happily hallucinate lyrics into silence; the gate is what prevents that.
Model download stalls. Hugging Face rate-limits occasionally. Delete the partial download in Settings → Storage & privacy, then retry.
Issues and pull requests are welcome — see CONTRIBUTING.md. Good first contributions:
- Better VAD (a Silero ONNX gate would beat the energy heuristic)
- Per-site insertion adapters for editors that resist
insertText - Translation of the UI strings
- Streaming decode that emits words while you're still speaking, not on pause
- Distil-Whisper and Whisper-Turbo support
Built on Transformers.js by Hugging Face, ONNX Runtime Web by Microsoft, and Whisper by OpenAI. ONNX conversions from the onnx-community organisation.
MIT © LocalDictate contributors.