Speak.We'llwriteitproperly.
Hold a hotkey, talk naturally, and what gets pasted reads like you typed it. Filler words gone. Punctuation correct. Formatting handled. Right where your cursor was.
MIT licensed · Open source · macOS 13+
This is the whole interface.
240ms
First-token latency
Whisper large-v3, streamed
3×
Faster than typing
~150 wpm spoken vs ~45 typed
100%
On-device storage
Plain JSON, deletable
One voice. Every app.
VoiceFlow pastes through the macOS Accessibility API, so polished text lands wherever your caret is — no plugins, no integrations to install. Scroll to watch one dictation travel through your stack.
Demonstration: as you scroll, a spoken transcript is typed, cleaned of filler words, and pasted as polished text inside Slack, Linear, GitHub, and Mail.
Four moves. Five seconds.
A complete dictation is four discrete beats — trigger, capture, polish, paste — each tuned to disappear into your attention.
Press the key. Anywhere.
A global hotkey opens the overlay wherever your cursor is — no window switch, no menubar hunt.
Speak naturally.
AudioWorklet capture streams straight to Whisper large-v3. Half-sentences, asides, code names — all of it lands.
AI cleans it up.
Fillers stripped, punctuation fixed, lists and code identifiers formatted — informed by the app you're typing into.
Lands in place.
Pasted through the Accessibility API exactly where your caret was. No clipboard hijack, no focus theft.
“Five seconds. Zero context switches. The keyboard generation finally got an interface.”
Custom dictionary
Teach the polish your jargon — product names, acronyms, framework slang.
Two modes
Toggle for long thoughts. Hold for quick zaps. Two shortcuts, one muscle memory.
Local history
Every dictation saved to your Mac as plain JSON. Searchable. Deletable. Yours.
Multilingual
English by default. Auto-detect across 90+ languages — even code-mixed speech.
We made some choices on your behalf.
Most voice-to-text tools optimise for the company shipping them. VoiceFlow is built on a different bet: you should own the data and the code. Four decisions, each made deliberately, each open to scrutiny.
LOCAL-FIRST
№01
Your transcripts never leave your Mac.
History, dictionary, and settings live as plain JSON in your user data folder. Nothing is uploaded, indexed, or trained on. You can read them, back them up, or delete them with rm. There is no server-side database. There is no server.
NO ACCOUNTS
№02
No sign-up. No telemetry. No waiting room.
Download the app, grant microphone and accessibility permissions, and start dictating. There's no login wall, no usage analytics, no growth funnel. The only network request VoiceFlow ever makes is the transcription itself.
OPEN BY DEFAULT
№03
MIT licensed. Read every line.
The whole app — main process, renderer, IPC, transcription pipeline — is on GitHub. Audit it. Fork it. Ship a private build for your company. We'd rather earn your trust than ask for it.
macOS, PROPERLY
№04
Built around how Macs actually work.
Tray-first. Accessibility-based paste. JXA for app context. AudioWorklet capture. Native menu items. It feels native because it is — no Electron skin painted over a web app, no cross-platform compromises.
Stop typing.Start talking.
Download VoiceFlow for macOS. Setup takes about a minute. After that, you stop noticing the keyboard.
● MIT·Open source·macOS 13+