MicText

Hold a key. Speak. Text appears — and your audio never leaves your machine.

▶ Try it in your browser — works on your phone too.

MicText is an embeddable STT library for the web plus reference push-to-talk dictation clients for macOS and Windows, all running Whisper locally. An optional self-hosted fallback server covers devices too slow to run the model.

Component	What it is
web/	Browser library: Whisper in a Web Worker (WebGPU → WASM). Headless createTranscriber() core + <mictext-mic> element. npm i mictext
mac/	macOS dictation client: hold right-⌥, speak, release — text typed at the cursor (Hammerspoon + ffmpeg + whisper.cpp)
win/	Windows dictation client: hold Right Ctrl (AutoHotkey v2 + ffmpeg + whisper.cpp)
server/	Optional self-hosted fallback (FastAPI + faster-whisper) for the web library's explicit slowDevice: 'server' opt-in

Dictation behaviour

All three clients share the same behaviour, tuned per platform:

• Warming vs. recording. Opening the capture device takes up to a second on the first press of a session. Each client shows that as a distinct warming state — a dim pill and a 🎙… icon on macOS, a grey pill and "warming up" tooltip on Windows, a warming attribute on <mictext-mic> — and only switches to the recording look once audio is genuinely being captured. Wait for it and no first words are lost.
• Silence is silence. A hold that carried no speech types nothing and shows nothing. Whisper's silence hallucinations ([BLANK_AUDIO], Thank you.) are dropped too.
• It learns your words. Correct a transcript (⌥⇧F on macOS, Alt+Shift+F on Windows, mic.learn() on the web) and the pair is remembered in ~/.mictext/terms.json — fed back to Whisper as decoding context and applied as a replacement. Local file, local only; the desktop clients share one file format.

Zero-upload guarantee

With the default configuration, no code path in this repo POSTs your audio anywhere. The web library only GETs model files (verified end-to-end with full network logging — see web/README.md); the desktop clients work with Wi-Fi off. The single exception is the web library's slowDevice: 'server' opt-in, which sends audio to a fallback URL you configure — typically your own server/ deployment.

Quick start

Browser — npm i mictext, then:

import { createTranscriber } from 'mictext'
const t = createTranscriber()
const { text } = await t.transcribeBlob(audioBlob)

macOS — one line (then hold right-⌥ and speak):

curl -fsSL https://raw.githubusercontent.com/dieterstemmet/mictext/master/mac/install.sh | bash

Windows — one line in PowerShell (then hold Right Ctrl and speak):

irm https://raw.githubusercontent.com/dieterstemmet/mictext/master/win/install.ps1 | iex

Configure mic, hotkey, and model from the tray Settings... menu, or win\config.ps1 from a clone.

(mac/README.md / win/README.md)

On your phone

Mobile browsers don't allow system-wide dictation, so there's no push-to-talk client for iOS/Android — but the web library runs fine on phones (the default whisper-tiny.en model exists exactly for mobile WebKit's memory limits):

• Use the hosted demo: open it on your phone, hold the mic, speak, copy the text anywhere. "Add to Home Screen" gives it an app icon.
• Embed it in any web app with <mictext-mic> — see web/.
• System-wide mobile dictation would need a native keyboard app (iOS keyboard extension / Android IME) — a possible future direction.

Design: headless core, swappable frontends

The durable interface is the headless transcriber (createTranscriber().transcribeBlob(blob) on the web; whisper-cli behind a hotkey on desktop). Dictation — "type what I said at the cursor" — is just the first frontend. The same ears can feed a voice assistant: that's the direction this project grows in, and why no client logic lives in the core.

License

MIT