What Laberampel measures, what the numbers can and cannot tell you, and why it is built the way it is.
Every voice gets two figures: its share of all speaking time, and its number of turns. Share answers "who filled the room". Turns answer "who got to speak at all".
The research points at turns. When Anita Woolley and colleagues measured what made groups good at solving problems together (Science, 2010), the strongest signal was how evenly people took turns, not how much anyone said. When Rachel Romeo and colleagues counted what happened at home between small children and adults (Psychological Science, 2018), the number of back-and-forth turns predicted the children's language, and the raw number of words did not. A four-year-old's fair share of a dinner conversation is not a quarter of the seconds. It is a turn whenever they want one.
Share still matters, because a room where one person takes every second turn but talks for a minute each time is not balanced either. So both are shown, and the family, the group or the class decides which "fair" it means tonight.
A turn begins when a voice starts speaking after someone else, or after it has itself been quiet for two seconds. It ends when another voice speaks. Under each name you see how many turns that voice took and how long the longest one lasted.
A short "mhm" from someone else ends a turn, so a person who is interrupted a lot will show more, shorter turns. That is a feature: it is what interruption looks like in numbers.
If someone who has already spoken has had no turn for five minutes while others kept talking, the app says so on their row. It is a hint to the adult at the table, never a verdict, and it never singles out the person who talks most.
When "Show loudness" is on, a row can say that a voice is louder or quieter than the group, by a few decibels. Treat this loosely. Phones and laptops adjust their microphone gain on their own, and someone sitting closer to the device sounds louder without speaking louder. Differences under three decibels are not shown at all.
The level-only mode is different: it measures the loudness of the room and nothing else. No voices are told apart. It is the app's version of the classic classroom noise light, with a line you set yourself by tapping "this is too loud" while the room is at that level.
Telling voices apart from one microphone on a table is approximate, and the app says so on every screen. Measured on public meeting recordings (a real table array with five to seven people, and replayed speech with eight), each voice's share is typically off by two to five percentage points, and in a large room by up to nine. The first minute of any session is unreliable while the app is still learning the voices.
Two things move the number more than any setting: distance (put the device in the middle of the table) and age. Children's voices are harder to tell apart: two unrelated ten-year-olds are counted as one person about three times as often as two adults. The "Children" preset raises the bar for joining an existing voice, which helps. Saying your name before the session helps more, because the app then knows whom to expect instead of guessing. Children under about seven are not covered by any recording set we could measure on, so treat their rows as rough.
These are numbers about the group, not evidence about a person. When a row looks wrong, it probably is; rename it, redo the enrolment, or laugh about it.
Put the phone in the middle of the table, tap a colour each, say your names, and eat. Look at the ring now and then. The point is not to hit equal numbers; the point is that the person who always talks notices, and the one who never gets a word in sees that someone noticed too.
The app fits the rituals that already exist: the Morgenkreis and the Kinderkonferenz in the Kita, the Klassenrat in primary school. It is the tool for the child whose job is to watch the time.
The adults vs children mode is meant for educators who want to see their own talk share. It does not tell children apart from each other and attaches no identity to anyone; it only sorts each stretch of speech into "adult" or "child". Decades of classroom research say that adults talk most of the time and that children learn more when they do not. Mary Budd Rowe found in the 1970s that teachers wait about a second after asking a question, and that waiting three seconds makes children's answers several times longer. A future version will show that wait time.
Known weakness, measured: the adult/child sorting is trained on voices aged six to forty-three and works well for children under twelve and adults over twenty-five, but it calls young women in their early twenties "child" almost half the time. In a Kita team that is many of the staff. Until this improves, read the adults-vs-children numbers as a trend across a session, not as a count, and expect the children's bar to be too high when the educator is a young woman.
A note on the law. Measuring an educator's own talk share is, technically, a device that monitors an employee's work. In Germany that falls under the works council's right of co-determination (§ 87 (1) no. 6 BetrVG) and, for church-run Träger, the equivalent rules of the MAVO. Laberampel is built so that this stays a matter of personal reflection: it keeps no history, exports nothing, shows results only on the screen of the person who started it, and is meant to be used voluntarily, by the educator, for themselves. A Träger that wants to use it should say so in writing, agree it with the staff representation, and never ask to see anyone's numbers. Children are not identified in these modes; no voice is stored.
Laberampel has no permission to use the internet, so nothing can leave the device. Audio is never written anywhere; it is processed as it arrives and discarded. What you see on the screen is the only place the information exists, and it is gone when the app is closed. The one exception, on Android only, is "Remember this speaker", which you have to switch on per person, with their agreement; it is stored encrypted on the device and deleted after five days without hearing that voice. The app recommends remembering only yourself and never a child.