Write at the speed you talk.

Double-tap a key, talk, tap again. The words land at your cursor, in any app, transcribed on your own machine.

macOS, paste into Terminal

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/ZyxWorks/murmurflow/main/install.sh)"

Windows, paste into PowerShell

irm https://raw.githubusercontent.com/ZyxWorks/murmurflow/main/install.ps1 | iex

View the source on GitHub

Paste it into Terminal on a Mac, PowerShell on Windows. Free, MIT licence. Read the script first.

On a Mac, a small waveform sits at the bottom of the screen while the microphone is open. It never takes the focus.

How it works

One line to install, two switches on a Mac. After that it is one key.

Paste the line
It installs what is missing, downloads the speech model once (about 1.6 GB) and starts with every login. Nothing needs to be there first. On Windows no Administrator either; on a Mac, Homebrew may ask for your password. A Mac then asks once for Microphone and Accessibility; you flip both switches.
Double-tap Control and talk
The microphone opens on the first tap, so your first word is not lost. Nothing is typed while you talk.
Tap again
The whole sentence lands at your cursor about two seconds later, in one paste. What was on your clipboard is put back.

Why bother

A 200-word message is five minutes of typing. Said, it takes eighty seconds, and your hands stay on the desk.

You speak150 wpmWithout trying

You type40 wpmAn average typist

It costs0 €MIT licence, no paid tier

Conversational speech runs about 130 to 150 words a minute.

What you are not paying for

No account
No subscription
No cloud server: whisper.cpp runs on your machine

Measured, not promised

One Mac: M4 Pro, macOS 26, the model large-v3-turbo, an 11 second sentence. Your numbers move with your machine.

Step	Seconds
Transcribe, warm, language on auto	1.9
Transcribe, warm, language set	1.2
First transcription after boot	13.3
Microphone open, the very first time	9.9
Speech lost at the start of a clip	0.04

The speech server starts at login and stays warm, so the slow first transcription happens while nobody waits.

Privacy

Your machine transcribes your voice, and the recording is deleted the moment that is done, before the text is pasted. The microphone closes with the key, and by itself after 15 seconds of quiet.

The text is not stored or sent anywhere. It goes through your clipboard once, and what was there before is put back. The only network request is the one-time model download.

Two opt-in settings change that, and both are off by default: polish runs a command you choose, which can be a hosted service, and keepAudio keeps the last clip and writes the transcript into the log, for debugging.

Every privacy detail, in the README

Questions

Does it work on Linux?

No. Wayland has no global hotkey API, so a key that works in every app is not possible there. murmurflow doctor runs on Linux and says so.

Is the microphone always on?

No. It opens on the first tap and closes on the last. A clip you forget closes itself after 15 seconds of quiet and pastes what you said.

Does it use an AI service?

No. whisper.cpp transcribes on your machine. You can add a polish step that runs a command of your choice, local or hosted. It is off by default, and it is the only way your words can leave the machine.

Does it change what I said?

No, with one exception: a closing "Thank you.", "Thanks." or "Danke." after a finished sentence is dropped, because whisper invents them in silence. If you really said it, it goes too. Nothing else is removed, reworded or reordered.

What is different on Windows?

There is no permission to grant, the fn key cannot be a trigger, and the clipboard is restored as text only. Typing into a window that runs as Administrator needs MurmurFlow to run as Administrator too.

Can my own voice agent share the key?

Yes. murmurflow pause --who "my agent" lends the key for a while, and the pause ends by itself. While dictation is on, the warm whisper.cpp server answers on 127.0.0.1:8479.

How do I stop or remove it?

murmurflow off turns dictation off until you turn it on again. murmurflow uninstall removes it.

Who builds this

ZyxWorks is a product studio. We built MurmurFlow for our own work and gave it away. The paid work is putting the same engineering inside a company.

What we do for companies