LocalDictate — Private Offline Voice Typing

Voice typing that never leaves your device. Whisper runs in your browser — press a hotkey, speak, and text lands in any field.

Speak freely. Type instantly. 100% private and offline. LocalDictate turns your voice into text in any text field on any website — Gmail, Notion, Slack, GitHub, Jira, a search box, an old web form. Put your caret where you want the words, press Ctrl+Shift+D, and talk. The part that makes it different: the speech recognition runs inside your browser. OpenAI's Whisper model is downloaded once and then executes on your own hardware through WebGPU. Your microphone audio is never uploaded, because there is nowhere to upload it to. There is no server, no account and no subscription. WHY THIS INSTEAD OF THE OTHERS Most dictation extensions stream your microphone to a company's servers. That means your half-finished emails, your medical notes, your client work and your private messages are transcribed on someone else's computer, under someone else's retention policy. LocalDictate has no backend at all. Turn off your Wi-Fi after the first model download and it keeps working perfectly. The entire source is on GitHub under the MIT licence, so you do not have to take our word for any of this. FEATURES • Whisper on your device — Tiny, Base and Small models, optimised for the browser • WebGPU acceleration with an automatic WebAssembly fallback, and it tells you which one is running • Near real-time: it transcribes as you pause, with a live preview while you speak • Types straight into inputs, textareas and rich editors; copies to the clipboard when a site uses a canvas editor • Voice activity detection so silence is never transcribed and never costs you compute • Automatic clean-up: fillers like "um" and "uh" removed, sentences capitalised, punctuation spaced • Spoken commands: say "new paragraph", "comma", "question mark" • Custom replacements for names and acronyms the model keeps mishearing • Hold-to-talk or press-to-toggle, with rebindable keys • Automatic language detection, plus 28 languages you can pin explicitly • Local transcript history you can export or switch off entirely • Dark and light themes • Free, open source, MIT licensed, no ads, no upsell HOW TO START 1. Install and grant microphone access — Chrome asks once 2. Choose a model. Base is the right first choice 3. Wait for the one-time download 4. Click into any text field and press Ctrl+Shift+D After that you are offline-capable forever. PRIVACY IN ONE PARAGRAPH LocalDictate collects nothing. No audio, no transcripts, no usage statistics, no crash reports, no identifiers. The only network request it ever makes is downloading the Whisper model from huggingface.co the first time you select it. Transcript history, if you leave it on, is stored in your browser profile on your computer and can be deleted with one click. Full policy: https://github.com/wushu75/LocalDictate/blob/main/PRIVACY.md REQUIREMENTS Chrome 116 or newer. WebGPU is strongly recommended for the larger models but is not required — LocalDictate falls back to WebAssembly on the CPU and still works well with Tiny and Base. The first download is 78 MB to 490 MB depending on the model you choose. KNOWN LIMITATIONS Google Docs draws text on a canvas with no editable DOM, so no extension can type into it directly. LocalDictate copies the text to your clipboard and prompts you to paste. Everything else that behaves like a normal text field works. OPEN SOURCE Source, issues and roadmap: https://github.com/wushu75/LocalDictate Licence: MIT ``` This submission (v1.0.4) fixes the "Load Model" functionality cited in the previous rejection (ref: Red Potassium). Root cause: the extension's speech-recognition worker failed to instantiate because the bundled machine-learning library (Transformers.js) was the wrong build variant — one that expected a build tool to resolve an internal dependency (onnxruntime-web). In a browser extension worker that dependency was unresolved, so the worker crashed immediately and no model could load. The package now bundles the fully self-contained build of the library, with the ONNX Runtime compiled in, so nothing needs to be resolved at runtime. To verify "Load Model": 1. Open the extension popup, or the options page (Engine tab). 2. Select the "Tiny" model (smallest, ~78 MB) and click "Download & load model". 3. A progress bar fills as the model downloads once from huggingface.co, then the status changes to "Ready" and the backend chip shows "WebGPU" or "WASM". 4. Click into any text field on any page, press the toolbar icon or hotkey, and speak. The model download is a one-time fetch from huggingface.co (declared in host_permissions and permitted by connect-src). After it completes the extension works fully offline. No audio or transcript ever leaves the device; there is no backend server. Full source, including the build fix: https://github.com/wushu75/LocalDictate