On-device dictation for macOS and iOS · powered by NVIDIA Parakeet
murmr is dictation that actually sounds like you wrote it — filler stripped, punctuation fixed, tone matched to the app. It all runs locally on your Mac and iPhone, so your voice never leaves your device.
One app for macOS and iOS · Pro is free for builders, or $149 once · voice and transcripts never leave the device.
Comment or tag @murmrvoice with your thoughts about murmr — we'll add them here.
There is no step four. That's kind of the whole pitch.
On Mac, press and hold the fn / 🌐 key in any app (⌥Space on the App Store build). On iPhone, tap the mic on the murmr keyboard and the app opens already listening. Double-tap to go hands-free.
Ramble. Backtrack. Say “um” a hundred times. murmr is listening on-device and isn't judging — that's the cleanup's job.
On Mac, let go — polished text lands at your cursor on the direct build, or sits copied and ready for ⌘V on the App Store build. On iPhone, swipe back and tap once. No cloud round-trip anywhere.
A live demo of the cleanup — the same filler-stripping, tone-matching pass the app runs on-device.
A menu-bar app on the Mac, a keyboard on iPhone — and a full voice toolkit behind both. Same engine, same privacy, no cloud in sight.
Apple doesn't let any keyboard touch the microphone — so murmr doesn't pretend. Tap the mic and the murmr app opens already listening. Say your piece, swipe back, and “Tap to insert” is waiting in the keyboard — filler stripped, punctuation fixed, tone matched. One tap drops your words at the cursor. That's the whole trip; anyone promising fewer steps is recording you somewhere else.
murmr lives quietly in the menu bar. Press and hold the fn / 🌐 key from any app and a floating pill shows it's listening; let go and the cleaned text appears where your cursor was. Double-tap to go hands-free.
murmr
Pro · 12dHold fn to dictate · double-tap for hands-free
“Let's ship the build tonight if that works.”
Behind the menu bar there's a full app: your dictation history with word counts and streaks, Meetings, Listen, the Assistant, Insights, your dictionary, snippets, style, and transforms — every tool in one sidebar, everything stored on this Mac. This is the actual window, not a mockup.
Dictation is half the product. Paste anything — a memo, a doc, that forty-minute read — pick an on-device voice, and press play. The system voice is free; on Pro the voices have names — Heart, Bella, Michael, and Emma on Kokoro, Nova and Atlas leading Supertonic's ten styles — and Marvis, new this release, speaks with real expression and clones your voice from a 10-second clip. Every model downloads once and runs on the Neural Engine, so nothing you paste goes anywhere either. On the Mac, select text in any app and have it read aloud — that's the floating Listen island below, mid-playback.
Hit record. When you stop, murmr transcribes the audio and separates the voices — speaker by speaker, time-stamped, entirely on your device. Rename “Speaker 1” to a real name and it sticks. There's no notetaker joining your call and no audio leaving the room: murmr transcribes what your mic hears.
Let's lock the launch date before we leave this call.
Marketing wants the 14th. The build is ready either way.
The 14th it is — I'll send the recap.
Voice memos, lectures, downloaded audio, even video — pick a file in the iPhone app and murmr returns a clean, time-stamped transcript without uploading a byte. Free handles clips up to five minutes; Pro removes the ceiling.
Transcribed on-device — nothing uploaded
Quick sync on the launch list.
The keyboard fix ships today; the meetings view starts tomorrow.
Push the demo to Thursday and we're clear.
No setup ritual. Open murmr, tap the mic, and talk. This is the actual iPhone Home screen — dictation, meeting capture, your word and streak stats, and voice insights, one tap each. The same on-device engine that powers the keyboard, in a tap-to-talk surface for longer thoughts.
Pick your default sound on first run — Standard, Formal, or Casual — and change it anytime in Style. Then teach murmr the names and jargon only your team says out loud, or let it learn them: correct a transcript and it offers the fix as a dictionary entry, never silently. Save snippets that expand when you say them. Build transforms — reusable rewrites that run on a hotkey or after every dictation.
Same on-device engine everywhere — what changes is what Apple lets each surface do. We'd rather tell you than let a permission dialog do it.
The build that does everything. Hold the fn / 🌐 key in any app, talk, let go — cleaned text auto-inserts right where you were typing.
Asks for three permissions — Microphone, Accessibility, Input Monitoring — because typing at your cursor genuinely requires them. Notarized download.
Fully sandboxed. Hold ⌥Space, talk, and your cleaned text is on the clipboard, one ⌘V from anywhere. That's the designed flow, not a fallback.
Apple's sandbox can't see the fn key, so ⌥Space is the default — rebind it to any combo with a modifier. The microphone is the only permission it ever asks for.
iOS bars every keyboard from the microphone, so murmr doesn't pretend. Tap the mic on the keyboard and the app opens already listening; swipe back and your cleaned text is one tap from the cursor.
Start from the keyboard, from Siri, or from the Action Button — the dictation itself is the same on-device engine everywhere.
Both Mac builds are the same app behind the trigger — menu bar, Listen, Meetings, transforms, the lot. Pick by how much you want murmr to type for you.
Windows · BetaA Windows build is in early Beta — the same on-device dictation, not yet at full Mac and iPhone parity. Download it and help us test. Android is on the way.
Not just speech-to-text. Speech-to-the-thing-you-actually-meant.
Transcription and cleanup both run on your Mac and iPhone. No servers, no audio uploads, no per-word fee. Your audio simply has nowhere else to go.
Every “um,” “uh,” and “you know” gets quietly deleted. Capitalization, commas, and sentence breaks land where they should — automatically.
Formal in Mail, punchy in Slack, tidy in Notes. murmr adapts the tone to whatever app you're typing into, so you don't have to.
Teammate names, product codenames, that one acronym only your team uses — add them once to your dictionary and murmr nails the spelling every time.
Mail, Slack, Notion, the address bar, a random text field at 2am. If you can type into it, you can talk into it.
A menu-bar app for the Mac — App Store or notarized direct download — and a custom dictation keyboard for iPhone. One voice, both pockets, zero cloud.
Ask a question out loud, get the answer spoken back — on-device via Apple Intelligence (macOS 26 / iOS 26).
One switch moves dictation to Parakeet v3 and its 25 European languages. Still entirely on-device.
Your recent dictations (the last 200) and every note are searchable. Keyword search is free; semantic search is Pro.
Double-tap and keep talking — a 10-minute safety timer finishes for you. Esc cancels instantly; your last finished dictation stays recoverable from the menu.
“Murmr quick dictation” opens the app already listening. Bind it to the Action Button and go.
Correct a word in a transcript once and murmr offers the fix as a dictionary entry. Suggested, never silent.
Streaks, peak hour, most-used word — computed locally, with a shareable stat card.
Dictate a quick note before the thought escapes. Notes stay on this device — by design.
A reverse trial, in plain sight — onboarding says so, the menu counts it down. When it ends, murmr degrades to free. And free is the product, not a demo.
If Pro lapses you get a one-time notice and the free tier — never a wall in front of your own dictation. And Pro itself is Free or $149: three builder steps unlock it for life, or one payment skips the ladder. See pricing and the full feature matrix →
murmr ships an MCP server that works with Claude Code & Codex out of the box — and any other MCP client. Your agent reads your transcripts and answers out loud, in an on-device voice.
Settings → Connections generates the config per client — Claude, Cursor, or a generic agent — and your builder token scopes every call to your own data.
{
"mcpServers": {
"murmur": {
"command": "node",
"args": ["…/murmur-mcp-server/index.mjs"],
"env": {
"MURMUR_BUILDER_ID": "…",
"MURMUR_API_TOKEN": "…"
}
}
}
}Copied from Settings → Connections — paste it, and your agent can talk.
Most “AI” dictation apps stream your microphone — and sometimes whoever you're on a call with — to a rented GPU in someone else's data center. That round-trip is their entire business model.
murmr deletes the round-trip. Speech recognition runs on the Apple Neural Engine; cleanup runs on Apple's on-device models. Your audio and transcripts have no upload path, because there's no server in the dictation loop.
Your voice takes a field trip. You pay per word.
Nothing leaves. Nothing to bill. Nothing to leak.
No account? Nothing syncs — there's nothing to send. Sign in, and exactly one kind of data leaves the device: settings. Uploading them is part of Pro; free devices only pull — which is how a purchase made on your Mac shows up on your iPhone.
Syncs — only if you sign in
Never syncs — signed in or not
The full policy is a two-minute read: privacy policy →
murmr is on the App Store for both Mac and iPhone — live today. Grab it for the device you're on, or get the Mac app as a direct download.
Prefer a direct download? Get the notarized .dmg · all download options →
Pro is Free or $149 — unlock it free by climbing the Builder Ladder (four verified steps, all open today), or buy it once for lifetime. See pricing →
We're continuously improving murmr's on-device models and engines — a sub-100ms streaming dictation engine and our own murmrvoice cleanup model are in the works. Drop your email and we'll let you know the moment a new, faster model lands. No audio leaves your device, then or ever.