What the front page left out.
The thresholds, the models, every receipt line, and the privacy notes in full.
The front page keeps to the short version. This page keeps the rest.
How the counting works.
Orange is you. Teal is everyone else, as one group.
Four counts sit under it: last minute, last five, last thirty, whole session. A conversation that turned two minutes ago shows up in only one of them.
A word nobody could be sure about is not counted on either side.
When unsure, it says so: I CAN'T TELL WHO'S TALKING RIGHT NOW.
How the models work.
The math comes from two courses at the Wharton School.
The first is a marketing course: a model of one customer's normal buying rate. Talking fits the same model, so tonight is read against your own rooms.
The second, an accounting course, teaches the reading scores analysts run over annual reports. They work on speech too, so the app runs them on your own words.
So you get two readings, not one. How much you talked, and how you talked.
Every session is scored before the model learns from it.
A session qualifies at five minutes, 150 of your words, and enough signal confidence. Your usual appears after three, and a percentile, forecast or nudge after eight.
Wharton did not make this app and has not endorsed it. I sat in the classes, built this afterward, and any mistakes are mine.
How the tap decides.
When your five-minute share passes your usual, the Watch taps you. The tap is a haptic, and the Watch stays silent.
A nudge that announces itself tells the table your phone said shut up. That is the opposite of the point.
It starts at 60% of the last five minutes, and lets go at 55%. Louder levels need 65% then 70% held, and nothing fires under 80 words.
It taps at your own 75th percentile, then your own 90th. Which means your quiet night and somebody else's quiet night are different nights.
One catches you talking more than usual, the other everybody else going quiet. The all clear waits for both to come back down.
It disarms. A confident false alarm is worse than silence.
Every line on the receipt.
The counts, the wheel, one stamp, and a line saying what was not kept. Nothing is stored: built when you ask, forgotten when you close it.
- WORDS
- Yours.
- OTHER WORDS
- Everyone else's, as one group.
- UNCERTAIN
- Words the app would not guess at.
- TOTAL AIRTIME
- Your slice of all the words in the room.
- LONGEST RUN
- Your longest unbroken stretch.
- TURNS
- How many times the floor moved.
- YOU CUT IN
- You went first while someone was still going.
- YOU WERE CUT OFF
- Someone went first while you were still going.
- SILENCES FILLED
- Times you started next after a short gap.
One of six, decided by the airtime. Overridden when the evening sits inside your own normal range.
| Solo session | JUST YOU |
| At or below your own 75th percentile | WITHIN NORMAL RANGE |
| Under 50% | EXCELLENT RESTRAINT |
| 50 to 60% | CIVILIZED |
| 60 to 70% | A BIT MUCH |
| 70% and up | RECEIPT SAYS SHUT UP |
A night you dominated can still read as normal. A quiet night can still be your loudest ever.
Share it as text, as one image, or print it on 58mm thermal paper. On paper: 32 columns, and a wheel redrawn one bit per dot.
NO AUDIO OR TRANSCRIPT WAS SAVED. These numbers, and nothing else.
The receipt prints the counts, the wheel and one stamp. Behind it sit eight pages of measures, each named for what it is.
The session · what was counted
Turn taking · the floor, and who held it
Words, turns, your longest unbroken run. Times you cut in, and times you were cut off.
Against your usual · where tonight landed
Your slice of it · how much of the room was you
Loud room, or a lot of you · what made tonight big
Two dials: how much you said, and how much of the room you took. A big night can be either, and they are not the same problem.
How you spoke · tempo, pauses, fluency
How you put it · readability and word choice
Speech rate, articulation rate, pauses, runs, fillers, repairs, and how fast you jump in. Then three reading scores, a repetition measure, and how much was about you.
How sure · what the numbers rest on
How much of the evening it could place, and how confident the recogniser was.
One set of numbers, two ways to read them. The wording setting changes words and never a number.
| PLAIN | TECHNICAL |
|---|---|
| YOU | OWNER |
| EVERYONE ELSE | NON-OWNER |
| YOUR SHARE | AIRTIME SHARE |
| YOU CUT IN | OVERLAP ONSETS, OWNER-INITIATED |
| HOW MANY PEOPLE YOU TALK LIKE | DOMINANCE INDEX (D) |
| WHERE IT SITS | STANDING, LEAVE-ONE-OUT |
How many people you talk like: one is normal, two is two people's worth.
Measured, then forgotten. No words are kept.
This reads word choice, not character.
Privacy, all of it.
- No audio file is ever made. Frames live in memory in a bounded ring and are dropped.
- No transcript is kept. A word becomes a few numbers in the speech callback, then it is gone.
- Speech is read on your iPhone by a model on your iPhone. If that model is missing, the session refuses to start.
- Other people are one anonymous group, never named, never read, never profiled. The only thing kept about the room is a headcount, from zero to four.
- Reading scores and delivery measures run on your own speech only. There is no code path that reaches them with another person's word.
- Your voice profile, 128 numbers plus calibration statistics, stays in this device's Keychain. It cannot be played back as speech.
- The app stores one kind of record: the numbers of a session.
- Sync is off, and there is no default server. Turned on, it sends your endpoint nine numbers, two timestamps, two random identifiers.
- The microphone is active only while a session you started by hand is running. Sessions never start on their own, and stop after ten minutes of no voice.
Start stays locked behind a switch that reads “I told the table”. Under it, in quotes: “This counts my talk balance. It saves no audio or transcript.”
A privacy test that cannot fail is not a test. It asserts no string is reachable, and once caught one in a dictionary key.
It does not measure emotion, stress or personality. It never even calculates jitter, shimmer, formants or absolute loudness.
It does not score anybody who is not holding the phone.
It does not tell people apart across a room. At a distance it places 48.4% of your words, so it abstains.
It does not print a rate before it has enough behind it. It shows the raw count instead.
It is a mirror, not a coach. Shut Up jokes about your numbers, never about you.
userWords otherWords uncertainWords longestMonologueSeconds turnCount interruptionsStarted interruptionsReceived silencesFilled averageConfidence plus two timestamps and two random identifiers
Requirements.
iPhone and Apple Watch. iOS 26 and watchOS 26.
Start it by hand. Tell the table. Read the receipt.
$1.99 once. No subscription. No account. No sign in.