<insert ttile suggestions>

Intro

We got a kid last year, the little fellow has just turned one and most of it been a fun ride. Except sleep. He has been as hard to but down as a helium baloon, usually it takes a solid 30+ min of sothing to get him under.

You can imagine the anger and frustraion then, when a modified exhaust vechical passes by your apartment, full throttle, usually a lonely man trying to prove his manlyhood, only to wake the kid up again just past midnight, third time that night, and your knees and arms hurt and you are desperate for sleep yourself.

When we bought a downtown aparment, we of course knew there would be noise, its part of the price. But these frequent, full trottle, by-passes are more then just noise, they are physically hitting your soul to the degree you are interrupted in anything you do, it’s not just sleep. Why do we allow this? Now the law obviously prohibs it, even through multiple paragraphs, but it is obvisously not systematicly enforced by goverment? I suspect part of the problem is that is also rather hard to quantify the problem. If i even were to complain, i would need to provide some hard facts, and that is what i set out to do.

Glossary

• dB(A) — a decibel reading weighted to match what the human ear actually hears, which is far less sensitive to very low and very high frequencies than a microphone is. Every level in this article is dB(A). My microphone is calibrated with a fixed +21 dB offset, set against a phone sound-level meter that is accurate to about 2 dB, and I deliberately set it at the low end of that margin.
• SEL (also written L_AE) — Sound Exposure Level. All the A-weighted sound energy of a single event, squeezed into one second. It folds loudness and duration into one number, so a lorry that drones past for twenty seconds outranks a motorcycle that cracks past in three, even if they peak at the same level. It is the right number for ranking events against each other, and it is what aircraft-noise research uses.
• dB2sMax — my own term. The loudest two-second window inside an event, energy-averaged. See the Limitations section for why I use it instead of a peak reading.
• Lden — day–evening–night level, defined in ISO 1996 and the metric Norwegian noise maps and municipal noise regulation are built on. It is the energy average of a whole 24 hours, with a 5 dB penalty added to the evening (19–23) and a 10 dB penalty to the night (23–07), because the same noise is worse when people are trying to sleep. One number per location per year. Remember it: most of this article is an argument about what it cannot see.

Brief overview of the findings

Setup

Of what i had laying around, I put a microphone from the pandemic days on my apartments building’s pitched roof, about 20 meters from the closest road. I used my Google Pixle to calibrate the sensitivty of the microphone, as it is said to be acurate to around around 2 dB margin.

By feeling alone, I record all sounds that are above a threshold of 72 dB(A) for more then 2 seconds. It seem to fit the pattern of what feels bothersome and worse. My windows provide around 27 dB dampering, so a 72 dB(A) sustained noise is felt as 45 dB(A) in our bedroom that faces the closest road. WHO recommends 30 dB(A) in bedrooms during the night for a sleep of good quality.

I’m trying to not report on anything untruthfull, so I’ve put my measurements on the right side of causion; the calibration is on the lower end of the 2-3 margin and the 2 second duration criteria removes a lot of short and sharp noises.

This never triggers for normal cars, motorcycles, busses or trucks. Emergancy vechicals do trigger, of course, and there has been some trucks that i would also categorize as “normal” (as opose to anti-social behaviour¹), but these have been carrying very heavy loads. I’m trying to be on the concervetive side here, and events that trigger my thresholds are genuinly loud. And of those that passing the threshold by a margin are very loud, we are talking little ripples in your glass of water.

I started recording in April 2026 and ran the system until medio July. For this research, I’ve decied to look at the full months of May and June, cutting out April which included some bugs, and July, that is object to vacation patterns. Minus the outage described below, that leaves 57 days — 4,898,546 seconds of measured sound, 99.5% of the wall-clock time in the window.

In addition to the events, I’ve also recorded the avg dB(A) level for each second during the whole period.

Limitations

My setup is sensitive to rain, therefor I have remove all events that contain sound of rain, as it artificially increases the reported levels to wrongfull heights.

I would have liked to write that there was very little rain in the 57 days, but that is not what the data says. Rain was detected in 2,416 events on 28 of the 57 days — 49% of them — touching 133 separate clock-hours, or 9.7% of the period. It is not a rounding error and I do not treat it as one: every rain-classified event is excluded from every count in this article, and every Lden figure below is computed with the rain-polluted stretches substituted out. (The two raw day/night averages I quote at the top of the findings are the exception, and they say so — they are there to match the ambient lines on the event figures.) What is fortunate is not that it rained little, but that rain turns out to be cheap to remove: taking it out moves the whole-period Lden by 0.5 dB.

I lost four days of data — 22–25 May, from 22 May 2026, 20:15 to 25 May 2026, 07:16 — when a bird pulled out one of my cables, cut the power and I was out of town. Those four calendar days are dropped from every figure and every number here, which is why the period is 57 days and not 61.

I do not know the output quality of the higher levels of my microphone, so instead of reporting on a dBMax on events, I’ve decieded to find the 2 sec windows that contains the loudes part of the event and avg that, calling it dB2sMax.

Findings

Firstof, from my second-resolution noise readings i’ve calculated that the avg sound level during the day (07:00–19:00) is 67.4 dB(A) and by night (23:00–07:00) is 63.7 dB(A). Those are the ISO 1996 day and night hours, so that they line up with the Lden figure in the next sentence; measured instead over the hours my kid is asleep (19:00–07:00) the night level is 64.8 dB(A). This forms the baseline of my noise environment, and it is the raw one — rain included, so that it matches the ambient lines drawn on the event figures further down. Cleaned of rain it reads 67.0 and 62.9 dB(A).

If i calculate Lden on this data, substituting the minutes where i knew rain has poluted my data with avg numbers, i get an Lden of 70.5 dB, against 71.1 dB with the rain left in. Blanking every clock-hour that had any rain in it at all, rather than only the seconds that triggered, gives 70.4 dB — so how strictly I treat the rain is worth about a tenth of a decibel, and nothing more. Everything below uses the rain-substituted figure.

A week on average looks like this.

And a slightly different reading of the same data is this heatmap

Over a 57-days period my system have triggered 11,874 events. 909 of them — 7.7% — I have listened to and classified by hand. The remaining 10,965 are classified by a machine-learning model: a Audio Spectrogram Transformer pre-trained on Google’s AudioSet, whose output I use as the input features for a small per-category classifier trained on my own hand labels. Held out from training, it gets 84.4% of clips right, 77.0% when every category is weighted equally regardless of how rare it is. Validation is grouped by day, never by clip, because otherwise two recordings of the same car on the same night end up on both sides of the split and the score is a fiction.

Statens Vegvesen has a measuring point just down the road (https://trafikkdata.atlas.vegvesen.no/utforsk?from=2026-06-30&display=chart&datatype=weekVolume&trpids=62991V604777) , and through May and June it counted 4.2 million vechicals. In the same window, I’ve recorderd and classified

• 4,509 anti-social behavior vechicals that pass my 72 dB(A) 2 sec avgerage threshold — 79.1 a day
• 2,227 russ events, which get their own section at the end
• 793 emergency vehcials with siriens
• 598 construction events
• 568 events of honking
• 267 helicopters pass-bys²
• 261 uncategorised, 151 of people, 83 of animals and a single one of music

I am listing russ and construction rather than folding them into a tail, because both are bigger than the siren count and hiding them would be the sort of thing I would want to catch someone else doing. russ is seasonal and construction is a daytime phenomenon; neither behaves like traffic, and both are treated separately below.

During my kids sleeping hours 19:00–07:00 the distribtuion of event categories are

• 1,776 vechicals
• 1,006 russ
• 189 siriens
• 189 honking
• 131 helicopters
• 89 events of people, of what i’d say is mostly alcohol induced behaviour
• 13 construction — it stops when people go home, which is the whole point

Rain is excluded here as everywhere else; for the record it contributed 1,208 events in the same hours, which is why I keep saying it had to be taken out. What the list actually shows is that vehicles are not merely the largest category overall, they are the largest category at night, by a factor of nine over sirens.

Of the 4,509 vechicals, lets look at the distribtuon of their loudness.

The median pass-by reads 80.3 dB(A) SEL and 74.5 dB(A) dB2sMax. The tail is the part that matters: 626 of them (14%) hold above 80 dB(A) for a full two seconds, 196 hold above 85, and 54 hold above 90 — through a wall, at 20 metres, with the windows shut. The loudest sustained two seconds any single vehicle managed was 98.5 dB(A).

Event samples

I’d say the average event is a motorcycle or car with a modified exhaust system going full throttle out of the clostest roundabout in any direction. If it passes by hour apartment, it is load, and if not, its usually a more distant sound for us, but usually just as painful for someone else. Here are some samples that you can play, and you can view its spectogram. The spectrograms have annotated data of how that particular sound end up in a SEL reading, and it shows where the trigger points are and where the usual background noise.

A sample week

One week, at full resolution, so that the aggregate numbers above have something concrete under them. The week is 22 June to 28 June, and I picked it on one criterion decided before I looked at its vehicle counts: it is the week with the fewest rain and russ events of any week in the period — 10 of them across seven days — so it is the week where the fewest readings had to be thrown away or set aside. It is not a quiet week. It carries 576 vehicle pass-bys, 82.3 a day, of which 88 hold above 80 dB(A) for two seconds.

One caution, because I would notice it if someone else did this: this week is an illustration, not the basis of anything. Every number in this article, including the Lden calculation further down, is computed over all 57 days. Nothing is argued from the week I chose.

Birds eye

Spanning the whole period, here is a raster showing all the events that are within my kids sleeping hours. One caveat on reading it: this figure was drawn on an 18:00–06:00 night, an hour earlier at both ends than the 19:00–07:00 window every count in this article uses. The shape is the same; the totals will not line up exactly with the lists above.

The worst continuous three hours of the day are 19:00–22:00: 919 vehicle pass-bys across the period, 16.1 on an average day. That is one loud vehicle every 11.2 minutes, for three hours, every evening. The single worst hour is 21:00 with 323 pass-bys over the period. And it does not stop there — it thins out through the night rather than ending.

How the municipality measures noise

Municipal noise regulation in Norway runs on Lden, and Lden is an energy average of a whole day, reported as a single yearly figure. So the honest test of my complaint is to ask what Lden would say if the loud vehicles simply vanished. I can run that experiment: I have a level reading for every second of the period, and I know which seconds belong to which event. So I took every second that falls inside one of the 4,509 vehicle events, replaced it with the background level of that same hour on that same day, and recalculated.

Those 4,509 events occupy 65,694 seconds — 1.34% of the period. Deleting every one of them moves Lden from 70.54 dB to 70.26 dB — a change of 0.27 dB. The evening period moves most, at 0.48 dB, and the night least, at 0.19 dB.

A difference of 0.27 dB is not audible. It is well inside the margin of my own calibration. It would not change a colour on a noise map, would not cross any regulatory threshold, and would not appear at all in a yearly figure. Remove every single vehicle that has woken my son in two months, and the metric my municipality would judge my complaint by does not move.

Discussion

It is usually cars and motorcycles with modified exhaust systems accelerationg full throttle out of the clostest roundabout, or the same vehicals reving their engines when the cars in front is not moving fast enough for their liking.

There is a control group for this, and it was in my data the whole time. Emergency vehicles are loud on purpose and nobody disputes that they should be. But they are disciplined about it: they run the siren when they need it and switch to lights when they do not, and it shows.

Sirens peak in the early afternoon and fall to single digits per hour between two and five in the morning; 189 of their 793 events land inside my son’s sleeping hours. Loud private vehicles peak at 21:00 — the evening, not the rush hour — and 1,776 of their 4,509 events land inside those same hours. The people whose noise is legally protected turn it off at night. The people whose noise is illegal do not.

Bonus; russ party emoji

As if this was not enough, there is one source of noise that I’ve yet to talk about. Russ. A sesonal celebration amoung high school gradutions students. We did not know of this when we bought the apartment, but when we got the the keys from the previous landlords, they wished us good luck with the russ celebrations, carrying a sense of relief in their tone of voice as they had bought something new outside of town. As part of their celebration, they gather in the middle of the clostest roundabout and drink one sip of alchol for each car they can get to honk.

It is not spread across the period. 2,227 events landed on 32 of the 57 nights, nothing at all before 2 May, and the three worst nights alone account for 48% of the total. The worst was 2 June, with 407 events in one night. A typical active night is 8. This is also the best free sanity check on my classifier that I have: a model that spread russ evenly across May and June would be visibly broken, and this one does not.