Dictation as a component

WordInk adds push-to-talk dictation to any text field in your web app. Hold a button, speak, let go: the text lands at the cursor, and the field's native undo still works. Use your own Groq, OpenAI or Deepgram key through a small relay you host, or a local model that runs in the browser. No WordInk account, no WordInk servers, no telemetry. MIT licensed.

<textarea id="message"></textarea>
<wordink-mic for="message" endpoint="/api/wordink"></wordink-mic>

Try it

This demo runs Moonshine tiny in your browser, so it needs no key and your audio never leaves this tab. The first press downloads the model (about 32 MB on WASM, about 80 MB on WebGPU); after that it is cached.

Dictate here

Hold the button (or Alt+D) and speak.

What you get

One element or one hook. <wordink-mic> for any page, including plain HTML with no build step, or useDictation for React.
Your provider, your key. Groq (default), OpenAI or Deepgram, switched with one attribute. See the comparison.
Keys stay on your server. @wordink/server forwards Groq audio and mints short-lived OpenAI and Deepgram tokens. It refuses every request until you give it an authorize check.
Offline option. @wordink/local runs speech recognition in the browser.
Nothing phones home. Every data flow is listed on the privacy page.

Packages

Package	What it is
`@wordink/web`	The `<wordink-mic>` element, as an ES module and a single-file CDN build
`@wordink/react`	`useDictation` and a typed `<WordInkMic>` wrapper (React 19)
`@wordink/core`	The engine: a Rust core compiled to WebAssembly plus a small TypeScript host
`@wordink/local`	Offline speech recognition (Moonshine via transformers.js)
`@wordink/server`	The credential relay for Cloudflare Workers, Node or any Fetch API runtime