Get unstuck

The direct build listens for the fn / 🌐 key system-wide and types the result at your cursor. That takes three permissions, each with exactly one job. Onboarding walks through them with live checkmarks; here’s the manual version:

1. Microphone— System Settings → Privacy & Security → Microphone → enable murmr. Hears you while you dictate; audio is processed on the Mac and never uploaded.
2. Accessibility— System Settings → Privacy & Security → Accessibility → enable murmr. Inserts the finished text at your cursor in whatever app you’re in.
3. Input Monitoring— System Settings → Privacy & Security → Input Monitoring → enable murmr. Notices the fn key being held from any app.

One macOS conflict to clear: System Settings → Keyboard → “Press 🌐 key to” → Do Nothing, so macOS doesn’t pop its emoji panel over your dictation. Onboarding flags this when it applies. Then hold fn anywhere and talk — the shortcut is rebindable in murmr’s Settings.

The App Store build runs fully sandboxed and asks for exactly one permission: the microphone. No Accessibility, no Input Monitoring. Two deliberate differences from the direct build:

• The trigger is ⌥Space, not fn. The fn key can’t be expressed as a sandboxed global hotkey, so this build uses a normal shortcut. Rebind it in Settings to any combo that includes at least one modifier — a bare key would swallow plain typing everywhere.
• Results land on your clipboard. Dictate, then press ⌘V; the status pill reads “Copied — press ⌘V to paste.” That’s the designed flow for this build, not a degraded one — it’s what lets it ship with a single permission.

The gestures are identical: hold ⌥Space to talk, release to finish, double-tap for hands-free.

1. Install the murmr app and open it once.
2. Go to Settings → General → Keyboard → Keyboards → Add New Keyboard… → murmr.
3. Tap the murmr keyboard entry and turn on Allow Full Access (what that’s for is covered below).
4. In any app, hold the 🌐 globe key on the keyboard and pick murmr.

Apple bars every keyboard extension from the microphone — true of every dictation keyboard on the App Store. murmr makes the round-trip fast instead of pretending it doesn’t exist:

1. Tap the mic on the murmr keyboard. The murmr app opens already listening.
2. Talk. murmr transcribes and cleans up on the phone.
3. Swipe back to the app you were typing in.
4. The keyboard shows your transcript as a preview chip — tap it to insert. Insertion is never automatic; nothing types itself.

A fresh transcript is offered for 5 minutes; older ones stay one tap away under Insert last.

Full Access exists solely so the keyboard can read your finished transcript from the shared on-device container the murmr app writes into — that handoff is the whole mechanism. The keyboard makes zero network requests — transcripts move through the on-device App Group container only — and the transcript file is cleared after you insert it.

All synthesis runs on-device via Core ML. Once a model is downloaded it works offline, permanently.

Two kinds of one-time downloads, both plain model-file fetches:

• Speech recognition — NVIDIA Parakeet, the model behind dictation, downloads on your first dictation. Turning on multilingual recognition (25 European languages) fetches the multilingual variant and applies on relaunch.
• Listen voices — each premium voice downloads the first time you pick it, and murmr shows the size before fetching, with a progress bar — a multi-hundred-MB download is never hidden behind a spinner. System needs no download.

A model download carries no audio, transcripts, or account data — it’s a file coming down, nothing going up. Use Wi-Fi and leave a few GB free if you plan to audition every voice.

Open Listen and pick a model and voice from the voice picker. On the free tier you get the System model with one voice at normal speed; Pro unlocks all four models with their full voice lists — Kokoro's four, Supertonic's ten, Marvis — plus 0.5–3.0× playback and audio export.

• macOS 14 or later · iOS 17 or later.
• AI cleanup tones and Transforms use Apple’s on-device foundation models, which require macOS 26 / iOS 26 with Apple Intelligence. On older systems murmr applies a built-in rule-based cleanup instead — dictation works everywhere.
• Disk space for the models you choose — sizes are listed above and shown again before each download.

The app itself is small. Dictation needs one on-device speech model (~220–450 MB), downloaded on first use. Listen voices are optional — the System voice needs no download, and premium voices download only if you pick one (sizes shown first).

Your transcripts and notes are text only, and no audio is ever stored. A typical install lands in the few-hundred-MB range; auditioning every premium voice is where it grows.

• Hold — press and hold the trigger (fn on the direct build, ⌥Space on the App Store build), talk, release. Capture starts the instant the key goes down, so your first word is never lost.
• Hands-free — double-tap the trigger and talk without holding; tap once more to finish. Hands-free has a 10-minute safety cap: if you walk away, murmr finishes the dictation on its own instead of recording forever.

Press Escwhile recording to cancel: nothing is transcribed, nothing is inserted. The ✕ on the recording pill does the same. (On the App Store build, Esc works while a murmr window is focused — dictating into another app, use the pill's ✕.) A cancelled take is discarded — the audio is never transcribed — but Recover last dictation in the menu-bar menu brings back your most recent finished transcript if you need it again.

Free dictation is batch: murmr transcribes the whole clip the moment you release the key. Pro adds streaming: words appear live in the recording pill while you’re still talking — settled text solid, the still-deciding tail lighter. Both paths run the same Parakeet model with the same accuracy; there’s no extra download, only different latency.

The pipeline, in order, all on-device:

1. Transcription (NVIDIA Parakeet).
2. Your Dictionary corrections — names, brands, jargon.
3. Tone-aware cleanup — Raw, Standard, Email, or Message, including spoken self-corrections (“scratch that”).
4. Snippet expansion — spoken phrases swap in their saved text.
5. An optional auto-Transform, if you’ve set one.
6. Text lands at your cursor (direct build) or on the clipboard (App Store build).

On Mac you can pin tones per app — Slack stays punchy while Mail stays formal — in Style.

No fixed cutoff like some apps impose. On Mac, hold-to-talk runs as long as you hold the key; hands-free has a 10-minute safety stop.

On iPhone, dictating with the app open is effectively unlimited. There's an optional Auto-endtimer (Settings → Style) that's off by default, plus a 3-minute cap that only applies if a dictation keeps running while the app is in the background.

Only if you choose a mode that rephrases. Pick Raw for exact verbatim (zero cleanup), or Standard to remove filler and fix punctuation without rewording your meaning. (Email and Message intentionally rephrase for those formats.)

Even in Standard, a built-in faithfulness check rejects any on-device AI edit that would swap or drop your actual words and falls back to a lighter rules-only pass. Want no AI at all? Set Cleanup to Quick — deterministic rules only.

Yes — Transforms are your own named rewrite instructions (your prompt), run on-device. Apply one to a selection, bind it to ⌥1–9, or auto-run one after every dictation.

Today you create and edit Transforms in the Mac appand they sync to your iPhone; on iPhone you apply the ones you've made plus the built-ins. Per-app tone rules (Slack punchy, Mail formal) live on Mac; iPhone uses a global tone plus auto-infer-from-field.

No. Dictation, cleanup, Listen, Meetings — everything works signed out. Sync is inert until you sign in; nothing is uploaded before that, because there’s nowhere to upload it to.

When you sign in, the sync payload carries:

• your Dictionary, Snippets, and Transforms,
• per-app tone rules and your preferences (tones, hotkey choice, Listen model and voice, toggles),
• your Pro entitlement, so a purchase follows you.

It never carries audio, transcript history, scratchpad notes, or usage stats. That isn’t a policy promise — those fields don’t exist in the sync wire format, so the app physically cannot upload them. The full accounting is in the privacy policy.

One honest asymmetry: every signed-in device pulls — so a Pro purchase made on your Mac reaches your iPhone — but uploading settings changes is part of Pro.

Auth tokens are stored in the macOS / iOS Keychain. Sign out any time — signing out stops sync entirely and your local data stays put.

In-app, on either platform — no email required:

• Mac — murmr Settings → Account tab → Delete account… → confirm. The button appears once you’re signed in.
• iPhone — murmr → Settings tab → Style → the gear icon → Sync → Delete account… → confirm.

Deletion runs server-side immediately: your synced settings and the account itself are removed. Data on your devices stays there — delete the apps to remove it.

murmr runs a reverse trial: every Pro feature is unlocked for your first 14 days, automatically — no card, no account. It’s never silent: onboarding says so and the plan row counts down (“Pro trial · 9 days left”). When it ends you get a one-time notice and murmr settles into the free core. No hard wall, no hostage features — dictation stays free forever, and the speech model is kept warm so day 15 isn’t a cold start.

Free forever: on-device dictation on Mac and iPhone, the cleanup engine with all four tones (one global default), Dictionary, 5 Snippets, 3 Transforms, Scratchpad, 7-day Insights, Meetings recording with speaker labels, file transcription up to 5 minutes (iPhone), keyword search, and the System Listen voice.

Pro:instant streaming dictation, contextual smart formatting, per-app tone rules, unlimited snippets and transforms with ⌥1–9 hotkeys, the Assistant, settings-sync upload, semantic search, 90-day insights, file transcription past 5 minutes, every premium voice — Kokoro's four, Supertonic's ten styles, Marvis with voice cloning — 0.5–3.0× playback, and audio export. The full matrix is on the pricing page.

Two real paths. Earn it: the Builder Ladder — join the Builder Discord, open a free mailpipe.dev builder account, set up Workflow TV — totals 100 points and grants lifetime Pro. Honest status: the Discord invite is live today; rung verification is opening soon and shows exactly that until it ships. Or pay once: $149 for lifetime Pro — an in-app purchase on the App Store builds, and a web checkout coming to murmrvoice.com for the direct build. No subscriptions, no per-word fees.

macOS ties Accessibility and Input Monitoring grants to the app’s code signature, and an update can silently invalidate them — the checkbox still looks on, but key events stop arriving. The fix:

1. System Settings → Privacy & Security → Input Monitoring — toggle murmr off and back on (or remove it with − and re-add).
2. Repeat under Accessibility.
3. Quit and relaunch murmr.

Click the murmr icon in the menu bar once — opening the menu re-arms the hotkey listener after fresh permission grants. If fn still doesn’t respond, quit and relaunch murmr; permissions granted while an event tap is running sometimes need a restart to take effect.

Also set System Settings → Keyboard → “Press 🌐 key to” → Do Nothing, so macOS doesn’t pop its emoji panel over your dictation. And if you’re on the App Store build: fn isn’t the trigger there — hold ⌥Space.

Another app probably owns that shortcut. Rebind murmr’s trigger in Settings to any combo that includes at least one modifier — a modifier-less key can’t be used (it would swallow plain typing), and murmr falls back to ⌥Space if you pick one.

Check the microphone permission. Mac:System Settings → Privacy & Security → Microphone → enable murmr. iPhone: Settings → murmr → Microphone. Audio never leaves the device either way — the permission is purely so the app can listen while you dictate.

First-use downloads run from ~180 MB to ~800 MB (sizes above), so a slow connection can take several minutes. Check the connection, then pick the model again to retry — a model that finished downloading is cached on the device and is never re-fetched. Make sure the device has enough free disk space for the model you picked.

Open the murmr app at least once, then Settings → General → Keyboard → Keyboards → Add New Keyboard… → murmr. If tapping the mic opens nothing, the keyboard shows a retry chip; if transcripts never come back after dictating, check that Allow Full Access is on — the keyboard tells you when it’s missing, because without it the transcript can’t cross back from the app.

Swipe back to the app you were typing in — the preview chip appears there, and a fresh transcript is offered for 5 minutes. Past that window, tap Insert last to pull up the most recent result. Nothing inserts without your tap, by design.