Your voice is the interface. It should stay yours.
murmr is dictation, text-to-speech, meeting transcription, and a voice assistant for Mac, iPhone, and Windows (Beta) — all of it running on the device in your hand. This page is the why.
The premise
Talking is faster than typing and closer to how you actually think. For years the catch was that making speech useful meant shipping it to a data center: your voice, your meetings, your half-formed ideas, uploaded so a server somewhere could turn them into text.
That catch is gone. Apple silicon runs NVIDIA's Parakeet speech model in real time on the Neural Engine, and Apple's on-device foundation models can clean a raw transcript into something you'd actually send. murmr is built on the one decision that follows: everything in speech happens on the device. Transcription, cleanup, tone matching, reading aloud, meeting speaker labels, the assistant's answers — local, all of it. After the one-time model download, airplane mode changes nothing.
Privacy as architecture, not policy
Most products promise privacy in a document. murmr enforces it in the build. There is no code path that uploads your audio or transcripts — recordings are temporary files deleted after processing, and your transcript history never leaves the machine that made it.
The one thing that can sync — settings, if you sign in — is a typed payload that carries your dictionary, snippets, transforms, and preferences, plus your Pro status so a purchase follows you. That is the entire list. Transcripts, notes, and audio aren't excluded by a checkbox; they are absent from the wire format itself. The claim, stated precisely: your voice and transcripts never leave your device. Settings you can see sync, only if you ask.
We don't have to promise not to read your words. We never receive them. The complete what-syncs list is in the privacy policy.
Paper, not glass
The apps are designed like paper: warm cream surfaces, ink-dark text, one lavender accent, edges you can see. No frosted panels to peer through, no translucency theater — software you read like a page.
The same instinct runs through the behavior. The iPhone keyboard never auto-inserts a transcript; you tap to place it. The Mac App Store build says “Copied — press ⌘V to paste” because that is exactly what happened. And when the Pro trial ends, nothing locks — dictation is free forever, so the app degrades politely instead of taking hostages. Honest states, visible mechanics, no theater.
A small team that ships
murmr is a solo-founder American software company. There's no “our team of 50” page because there is no team of 50 — there's a founder, the shipped code, and a rule this site is written under: if a feature is described here, it exists in the app you can download. When something isn't built yet, we say so.
Small is the point. On-device software has no per-word cloud bill to recoup, which is why the free tier is real and permanent — and mail to support@murmrvoice.com reaches someone who can actually fix your problem.
Standing on open shoulders
Speech recognition is NVIDIA Parakeet (CC-BY-4.0) running through FluidAudio's CoreML runtime (Apache-2.0). Cleanup and answers come from Apple's on-device foundation models. The premium Listen voices — Kokoro, Supertonic, Marvis — are open neural TTS engines running on the Neural Engine. murmr is what happens when you take excellent open models seriously.