Murmur

Murmur is a dictation app that lives in the Mac menu bar. Hold a key, talk, let go, and cleaned-up text lands wherever the cursor already was. It all runs on the laptop: no account, and the audio and the transcript don't leave the machine.

I publish a lot with a model helping. The risk was never the tool. It's that the writing comes out sounding like every other model's, and people stop reading the second something feels machine-made.

Talking is the cheap half. While I dictate, it's counting how I write. Every dictation. Every correction on top of one. Every archive of old writing I pull in.

The counting is arithmetic, and no model gets a say in it. Murmur doesn't ask an assistant what my style is. It counts the same text the same way every time, and what comes out is a Voice Card: a short list of instructions any assistant can follow to write like me. It's fitted to one person's writing and I'm the only person in it.

That card is why the rest of the bench leans on this one app. This site won't publish a page until the card's scorer agrees the prose sounds like me. Distribution Engine puts every version it writes through the same check, Wellington Whales runs the card over a second site, and Meeting AI Assist carries it into a live call where the answers come out in my words. Same card in all of them.

The scorer runs out of room

I ran the scorer over the eleven pages already published here and all eleven came back at 100 out of 100. Sounds like you. There's nothing above that. Then I handed the same eleven pages to an outside reader that judges whether writing looks machine-made. One page per fresh conversation. It didn't call a single one machine-written. It called all eleven over-shaped, and it named the same few habits every time.

Both readings are correct, and that's the whole problem. My scorer asks whether a habit of mine is present at my rate. The reader asks how thick that habit runs against ordinary prose. Do those questions agree? They do, right up until the card starts working. Take a habit I genuinely have, push it to the top of its measured band, and my score sits pinned at its ceiling with nothing left in it to notice with.

So I built a fix and deleted it. My email lands on an ask about a fifth of the time and these pages landed on one zero times, so I added a floor: one section per piece had to close on an ask. Then I wrote four write-ups twice each from an identical brief (once with the floor, once without), and sent all eight back to the reader. The treated ones came back with a fresh complaint of their own. That's my fix showing up as the tell. It didn't last the day.

What I refused to fix

Some of what I measured I left alone. These pages use fewer sentence fragments than I do, and they vary their sentence lengths more than I do. Tightening either would've moved the card away from me, and I'd have called it a fix while I did it.

Where it stands: the layer that catches the over-shaping sits apart from the card, behind a switch, and the build I ship has it off. It's on in my own defaults. It goes on for everyone the day the drafts it produces stop sounding less like me than the ones written without it. I'm not there yet.