The sensor does not record audio, by design, so it cannot be a recording of anything. It reports a level and a frequency band, and between bookings it reports fifty eight decibels in the band a human voice sits in, in three second bursts.
I manage eleven holiday lets for other people. It is my whole job and the thing that can end it is a party, because a party in a residential street means a complaint, a council licence review and an owner who takes their house back.
So every one of them has a noise sensor.
I want to explain what those are, because the first thing anybody says is that I am recording my guests, and the design of the thing makes that impossible.
It is a small white disc on a ceiling. It has a microphone, and what it does with the microphone is measure. It works out a level in decibels and it works out roughly which frequency bands that level is in, and it sends those numbers to a server every minute. It does not store audio, it does not transmit audio, and there is no way to listen to a room through it. That is not a privacy promise in a brochure, it is why they are legal to fit, and it is audited.
What I see is a graph. A level line, and a set of bands, and a threshold I set. Over the threshold for long enough, I get a message.
Number seven is a three bedroom semi in a quiet road. It is the best performing house on my books and it has never had a complaint.
Since the sensor went in, in 2021, it reports noise in that house between bookings.
Not a hum. Not traffic. Bursts: two to four seconds at a time, between fifty four and sixty one decibels, with gaps of a few seconds, for twenty minutes to an hour.
Fifty eight decibels at a ceiling sensor is somebody talking in the room at an ordinary volume.
The bands are the part that made me stop filing it as a fault.
The dashboard breaks the level into bands, and the reason it does is to tell a party from a washing machine. Bass heavy and constant is music or an appliance. Mid band, in bursts, with gaps, is people.
There is a number on the dashboard they call speech likelihood, and I do not know exactly how it is worked out and I am not going to pretend I do. Across eleven houses and four years it goes high when people are in a room talking and low when the house is empty and a boiler is running, and that is as much as I need from it.
In number seven, between bookings, it is as high as it gets.
The system does not know the house is empty. It has no idea about my bookings. It just says that the pattern at that ceiling is people talking.
What I did, in order.
Checked the bookings. Every one of the ninety or so events is in a gap with nobody in the house. I know that from the lock, which is a separate system on a separate account.
Checked the sensor. Swapped it for a new one of the same type. Then, in 2022, I had a different manufacturer's unit put in alongside it, because I was willing to pay for two dashboards if it settled the question.
They agree. Both of them, on the same nights, to within about two decibels.
Checked next door. This was the obvious answer for a year, because a semi shares a wall and voices come through. The sensor was moved to the back bedroom, which shares no wall with anybody, and the events carried on at the same level.
And I did something I am less comfortable about, which is that I asked the neighbours, politely, whether they had people staying, on four specific dates. Two of those dates they had been in Spain.
The time.
Every event starts between 21:40 and 22:20. Not one outside that.
They last between twenty minutes and an hour and eight, and the longer ones have a shape: the bursts get further apart towards the end, which is what a conversation does when it is winding down.
I have the graphs. I have sat with them next to graphs of actual guests in actual houses and they look the same. There is nothing about the traces that marks number seven out except that there is nobody in the building.
What is in the room.
I have been in, twice, during an event. I live nineteen minutes away.
The first time I let myself in at about ten past ten with the dashboard live on my phone showing fifty seven decibels, and the house was silent. Not quiet. Silent, the way an empty house is, with the fridge off and the heating off and a clock in the kitchen I could hear from the hall.
The dashboard, in my hand, standing in that hall, was reading fifty seven decibels, in bursts, and the speech number was high, and I could hear a clock.
Then it stopped, about four minutes after I came in, and the graph is flat from that point.
The second time I did not go in. I sat in the car outside with the window down for fifty minutes while the dashboard ran, and you cannot hear a normal conversation from the road in that house, because I have tested that with two of my own staff standing in the lounge talking, so the road tells me nothing.
What I did see is the hall light go on at about twenty five past ten and off again about a minute later, which is the motion sensor on the hall light, which I fitted, and which does not come on for nothing.
I have not told the owner and I am not going to, and my reason is not a kind one.
Number seven earns her about nineteen thousand a year and it earns me my commission on that, and she is not a woman who would hear any of this and shrug. She would have the sensor out, and then the house would have no sensor, and then one night there would be an actual party and a complaint and a licence review, and the house would come off the market, and it would be because of a thing I cannot even describe without sounding unwell.
So the threshold on number seven is set to sixty five decibels instead of fifty five, which means nothing it has ever done will alarm, and the events go into the history where I can see them and nobody else looks.
And once a month or so, at about ten at night, in a house with nobody in it and the heating off and the water off at the stopcock, a room full of people has a conversation that gradually runs out, and a ceiling disc that cannot hear words counts them anyway, and tells me how loud they were.