PrivacyFlow

Private dictation for the Mac.
Hold a key, speak, let go. Your words appear where you are typing, and they never leave your Mac.

Private by design

Many dictation tools send your voice to a server. PrivacyFlow never does.

• Everything runs on your Mac. Speech recognition and text cleanup happen on its own GPU.
• No account, no cloud, no telemetry. There is nothing to sign up for and nothing phones home.
• The network is used only to download the speech models, once, on first launch. Every file is checked against a fixed checksum.
• Recordings are deleted immediately. Audio exists only while it is being transcribed.
• Nothing is kept. Recent dictations are shown in the console while PrivacyFlow runs and are gone when you quit.
• The microphone is off between dictations. It listens only while you hold the key.
• Open source, so you do not have to take any of this on trust.

What it does

PrivacyFlow turns speech into finished text, not just a transcript. A small model reads each dictation and decides whether it can go in exactly as spoken, or needs tidying first. When it does, S1-mini by Superwhisper rewrites it: punctuation fixed, spoken corrections applied ("Friday, actually no, Thursday" becomes "Thursday"), and a spoken list turned into a list. Most dictations go in untouched.

It works in almost any app you can type in.

Fast

PrivacyFlow is built so your text appears the moment you let go of the key.

• Most dictations skip the language model. The small routing model reads each transcript first, and text that can go in as spoken goes straight in. Only dictations that need tidying wait for S1-mini.
• The models never go cold. They are locked in memory, so macOS cannot compress or swap them out while you are not dictating. Measured right after the Mac was put under memory pressure, routing took about 0.1 seconds and a rewrite about 0.15 seconds. Without the lock, the same steps took 2.7 to 6.7 seconds.
• No round trip. Everything runs on the Mac's GPU, so there is no server to wait for.

How it compares

All of these open-source apps can turn speech into text on your Mac. The difference is what happens to your words next. To tidy up a dictation or apply a spoken correction, the others use a language model, and every one of them can be set up to send your text to a cloud AI to do it. PrivacyFlow does all of it on your Mac, every time. It has no code that could send your words anywhere.

	PrivacyFlow	Handy	FluidVoice	OpenWhispr	VoiceInk	Whispering
Can send your voice to the cloud	Never	No	No	Yes, its cloud is the recommended setup	Optional	Optional
Tidies text with a language model	✅ On by default	Off by default, experimental	Off by default	✅ On by default	Off by default	Off by default
Can send your text to a cloud AI	Never	Optional	Optional	Yes, its cloud is the recommended setup	Optional	Optional
On-device cleanup with nothing else to install	✅	Apple Intelligence only, macOS 26	✅ Closed-source runtime	✅	✅ With 16 GB of memory	❌ Needs Ollama or similar
Skips the language model when the text is already fine	✅	❌ You choose with a second hotkey	❌	❌	❌	❌
Applies spoken corrections	✅	With your own prompt	✅ With cleanup on	✅	✅ With cleanup on	With your own prompt
Saves your dictations by default	No	The last 5, with audio	Transcripts	Transcripts, and audio for 30 days	Transcripts and audio	Transcripts and audio
Usage analytics	None	None	On by default	Opt-in	None	On by default
Price	Free	Free	Free	Free, paid cloud plans	$25 or more for the prebuilt app	Free

"Optional" means you can connect a cloud provider, usually with your own API key. Checked against each project's own source code and documentation on 29 September 2026: Handy 0.9.7, FluidVoice 1.6.9, OpenWhispr 1.10.2, VoiceInk 2.21 and Whispering 7.11.0. If something here is out of date, please open an issue.

Where the others are ahead: several run on Windows and Linux, and all of them understand many languages. PrivacyFlow runs only on Apple Silicon Macs and understands English.

How to install

You need a Mac with Apple Silicon (M1 or later) running macOS 14 Sonoma or later, and about 3 GB of free disk space.

1. Download PrivacyFlow-<version>.dmg from the latest release.
2. Open it and drag PrivacyFlow into Applications.
3. Open PrivacyFlow from Applications.

PrivacyFlow is not yet signed with an Apple Developer ID, so macOS blocks the first launch and says it cannot verify the app. To allow it, open System Settings → Privacy & Security, scroll down to the message about PrivacyFlow, and click Open Anyway. You only need to do this once for each version.

First launch

A setup window walks you through the rest:

1. Download the speech models, about 2.5 GB in all.
2. Grant three permissions in System Settings → Privacy & Security. The setup window has buttons that open each page.

Permission	Why PrivacyFlow needs it
Microphone	To hear you while you hold the dictation key
Accessibility	To paste the text into the app you are using
Input Monitoring	To notice the dictation key while another app is in front

If PrivacyFlow is missing from the Input Monitoring list, add it with the + button. Quit and reopen PrivacyFlow after granting Accessibility and Input Monitoring; macOS applies them at launch. macOS may offer to do this for you, and that is safe even while the models download: the download picks up where it left off.

How to use

Hold Right Option, speak, and let go. The capsule shows that PrivacyFlow is listening, then working, and the text appears at your cursor.

• Keep talking. You can start the next dictation while the last one is still being processed. They are inserted in the order you spoke.
• Nothing gets lost. The text is also left on the clipboard. If you switched apps before it was ready, it is not pasted into the wrong one; paste it yourself with ⌘V.
• Open the console from the menu bar icon, or by right-clicking the capsule, to see recent dictations, change settings, check status, or quit.

In Settings you can:

• Choose the dictation key. Right Option is the default. Any modifier works on its own (Left Control, Right Command, Fn and so on), as do function keys and combinations such as ⌃⌥ plus a letter. PrivacyFlow refuses keys you type with, and warns about keys macOS uses itself.
• Pick a microphone. PrivacyFlow uses the system input unless you choose one.
• Turn the sound cues off.
• Use minimal mode, which folds the capsule into a small button until you dictate.

PrivacyFlow understands English.

Troubleshooting

The dictation key does nothing while another app is in front. Input Monitoring is missing. Grant it, then quit and reopen PrivacyFlow.

The text is transcribed but never appears. Accessibility is missing. The text is on the clipboard in the meantime.

A permission is switched on but still does not work, usually after an update. Because PrivacyFlow is not yet signed with a Developer ID, macOS treats each new version as a different app. In System Settings → Privacy & Security, select PrivacyFlow in the list, remove it with the − button, and add /Applications/PrivacyFlow.app again with +. Switching it off and on again does not help.

The dictation key stops working in a password field or some terminals. That app has turned on secure keyboard entry, which hides the keyboard from every other app. The Status tab says when this is happening.

Uninstall

Quit PrivacyFlow, then delete:

• /Applications/PrivacyFlow.app
• ~/Library/Application Support/PrivacyFlow (the models and your settings)
• ~/Library/Caches/io.github.fz77z.privacyflow

You can also remove PrivacyFlow from the permission lists in System Settings → Privacy & Security.

Under the hood

The app is written in Swift. The models run in a Python process bundled inside the app, on the Mac's GPU through MLX, and stay locked in memory so the first dictation after a long break is as fast as any other. That uses about 2.5 GB of memory while PrivacyFlow runs.

Step	Model
Is this speech?	Silero VAD
What was said?	Whisper large-v3-turbo, 8-bit
Does it need a rewrite?	Kev router, built on Qwen2.5-0.5B
The rewrite	S1-mini by Superwhisper, 8-bit

See docs/architecture.md for how it fits together, and docs/development.md to build it yourself.

License

PrivacyFlow is released under the MIT License. The models and bundled libraries keep their own licenses, listed in THIRD_PARTY_NOTICES.md.

S1-mini is made by Superwhisper. Speech recognition is OpenAI's Whisper, run with Apple's MLX.