Murmur

Dictation for Mac that never sends your voice anywhere. Hold Right ⌘, talk, let go — the text is typed in at your cursor, in whatever app you were already using. Transcription runs entirely on your Mac.

Part of the Mac suite · in private beta — ask for a build

Apple silicon, macOS 15 or later. Signed and notarized by Apple; the app updates itself.

What it does

• Push to talk. Hold Right ⌘ to dictate, double-tap to lock it hands-free, and press any other key while holding to cancel — because you were probably just doing ⌘C.
• On-device transcription. NVIDIA's Parakeet model, running through CoreML. The model downloads once and then works offline, on a plane, in a basement.
• Warm start. The audio engine stays configured between dictations, so speaking starts in under 100 ms and idle Murmur measures 0% CPU.
• Pastes where you're typing. Your clipboard is snapshotted and restored, so dictating doesn't cost you whatever you had copied.
• Dictation modes. Raw by default. Switch to Email, Message, Bullets, Concise, Clean-up or Formal and the transcript is rewritten on-device by Apple Intelligence before it lands. Auto picks the mode for you based on the app you're in and how the sentence opens.
• Sounds like you. Paste in a few samples of your own writing and Murmur distills a style card on-device, so rewrites keep your voice instead of flattening it.
• Your vocabulary. Add names, jargon and product terms. Murmur biases the acoustic model toward them rather than blindly search-and-replacing, so a real English word is never quietly overwritten.
• Meeting notes. It notices when a call starts, records your mic and the system audio as separate tracks, and writes a markdown note with an on-device summary. Cancel and nothing is kept.
• Smart formatting, deterministically. Spoken emails and URLs become real ones, and numbers become digits where the context is numeric. Same input, same output — no model involved.
• Pauses your music. Whatever was playing stops for the length of the dictation and resumes after. It will never start audio that wasn't already playing.

Privacy

Audio and transcripts stay in ~/Library/Application Support/Murmur/. Transcription and every AI rewrite run on-device through CoreML and Apple Foundation Models — there is no transcription server and no account. The only outbound request is the daily update check.

Murmur needs Microphone access to hear you and Accessibility access to type at your cursor. System audio capture is only requested if you record meetings.