open-wispr

open-wispr.com
Local, private voice dictation for macOS. Hold a key, speak, release — your words appear at the cursor.
Everything runs on-device. No audio or text ever leaves your machine.

Powered by whisper.cpp with Metal acceleration on Apple Silicon.

Install

curl -fsSL https://raw.githubusercontent.com/human37/open-wispr/main/scripts/install.sh | bash

The script handles everything: installs via Homebrew, walks you through granting permissions, downloads the Whisper model, and starts the service. You'll see live feedback as each step completes.

A waveform icon appears in your menu bar when it's running.

The default hotkey is the Globe key (🌐, bottom-left). Hold it, speak, release.

Full installation guide — permissions walkthrough with screenshots, non-English macOS instructions, and troubleshooting.

Uninstall

curl -fsSL https://raw.githubusercontent.com/human37/open-wispr/main/scripts/uninstall.sh | bash

This stops the service, removes the formula, tap, config, models, app bundle, logs, and permissions.

Configuration

Edit ~/.config/open-wispr/config.json:

{
  "hotkey": { "keyCode": 63, "modifiers": [] },
  "modelSize": "base.en",
  "language": "en",
  "spokenPunctuation": false,
  "maxRecordings": 0,
  "toggleMode": false
}

Then restart: brew services restart open-wispr

Option	Default	Values
hotkey	63	Globe (63), Right Option (61), F5 (96), or any key code
modifiers	[]	"cmd", "ctrl", "shift", "opt" — combine for chords
modelSize	"base.en"	See model table below
language	"en"	"auto" for auto-detect, or any ISO 639-1 code — e.g. it, fr, de, es
spokenPunctuation	false	Say "comma", "period", etc. to insert punctuation instead of auto-punctuation
maxRecordings	0	Optionally store past recordings locally as .wav files for re-transcribing from the tray menu. 0 = nothing stored (default). Set 1-100 to keep that many recent recordings.
toggleMode	false	Press hotkey once to start recording, press again to stop. Default is hold-to-talk.

Models

Larger models are more accurate but slower and use more memory. The default base.en is a good balance for most users.

Model	Size	Speed	Accuracy	Best for
tiny.en	75 MB	Fastest	Lower	Quick notes, short phrases
base.en	142 MB	Fast	Good	Most users (default)
small.en	466 MB	Moderate	Better	Longer dictation, technical terms
medium.en	1.5 GB	Slower	Great	Maximum accuracy, complex speech
large-v3-turbo	1.6 GB	Moderate	Great	Fast multilingual, near-large accuracy
large	3 GB	Slowest	Best	Multilingual, highest accuracy (M1 Pro+ recommended)

Non-English languages: Models ending in .en are English-only. To use another language, switch to the equivalent model without the .en suffix (e.g. base.en → base) and set the language field to your language code. Multilingual models are slightly less accurate for English but support 99 languages.

If the Globe key opens the emoji picker: System Settings → Keyboard → "Press 🌐 key to" → "Do Nothing"

Menu bar

Click the waveform icon for status and options. Recent Recordings lists your last recordings; click one to re-transcribe and copy the result to the clipboard.

State	Icon
Idle	Waveform outline
Recording	Bouncing waveform
Transcribing	Wave dots
Downloading model	Progress ring
Waiting for permission	Lock

Click the menu bar icon to access Copy Last Dictation — recovers your most recent transcription if you dictated without a text field focused.

Compare

	open-wispr	VoiceInk	Wispr Flow	Superwhisper	Apple Dictation
Price	Free	$39.99	$15/mo	$8.49/mo	Free
Open source	MIT	GPLv3	No	No	No
100% on-device	Yes	Yes	No	Yes	Partial
Push-to-talk	Yes	Yes	Yes	Yes	No
AI features	No	AI assistant	AI rewriting	AI formatting	No
Account required	No	No	Yes	Yes	Apple ID

Privacy

open-wispr is completely local. Audio is recorded to a temp file, transcribed by whisper.cpp on your CPU/GPU, and the temp file is deleted. No network requests are made except to download the Whisper model on first run. Optionally, you can configure open-wispr to store a number of past recordings locally via the maxRecordings setting. Those recordings stay private and on your machine, and we default to not storing anything.

Roadmap

See what's planned and in progress on the public roadmap. Feature requests and ideas are welcome as issues.

Build from source

git clone https://github.com/human37/open-wispr.git
cd open-wispr
brew install whisper-cpp
swift build -c release
.build/release/open-wispr start

scripts/dev.sh — interactive dev build

Running bash scripts/dev.sh is the easiest way to configure, build, and launch a local development build in one step. It does the following:

1. Reads your current config from ~/.config/open-wispr/config.json and shows the existing values as defaults.
2. Interactively prompts you to update every setting:
   • Model size — choose from eight options (e.g. tiny.en, base.en, small.en, medium.en, or the equivalent multilingual variants).
   • Language — ISO 639-1 code (e.g. en, fr) or auto for auto-detect.
   • Spoken punctuation — whether saying "comma", "period", etc. inserts punctuation marks.
   • Max recordings — 0 = temp-only / nothing stored (privacy default); 1–100 = keep that many recent recordings locally for re-transcription.
   • Toggle mode — false = hold-to-talk (default); true = press once to start, press again to stop.
   • Hotkey — enter a human-readable key name such as fn, f5, or a chord like cmd+shift+r.
3. Writes the chosen settings to ~/.config/open-wispr/config.json.
4. Stops any running open-wispr start process and the Homebrew-managed service.
5. Removes the Homebrew-installed open-wispr formula (if present) so the locally-built binary takes precedence.
6. Installs whisper-cpp via Homebrew if it is not already present.
7. Builds the Swift app from source in release mode (swift build -c release).
8. Bundles the binary into OpenWispr.app via scripts/bundle-app.sh.
9. Copies OpenWispr.app to ~/Applications/ so macOS recognizes it for microphone / accessibility permission grants.
10. Launches ~/Applications/OpenWispr.app/Contents/MacOS/open-wispr start.

Support

open-wispr is free and always will be. If you find it useful, you can leave a tip.

License

MIT