A column about curious stories from the worlds of science and technology. Today — about a sound we have unlearned to hear. A plane passes outside your window: the brain tags the rumble as “background” and strikes it from consciousness in half a second. Recording equipment disagrees: buried in that background is a geographic fix accurate to within a few hundred meters. Not a metaphor, but a method that, in early November 2026, made the jump from academic journals to an ordinary user’s computer. And yes, turning off geotagging in your phone settings will not save you from it.
Let’s start with the case that swept through the tech feeds. An author going by the handle colsto posted a video with a deservedly long title: “I discovered a sound that leaks your exact location (and it's everywhere).” On November 2, it landed on Lobsters, and for part of the audience, the evening was productively ruined.
The task looked hopeless: a blurry recording from a surveillance camera, trees and a riverbank in frame, the only clues a timestamp reading “September 8” and the certainty that it was somewhere in the continental United States. No signs, no house numbers. On ordinary playback, the audio track says nothing either: quiet outdoor ambience, nothing meaningful. But on a spectrogram — a display of frequencies over time — two distinctive “accordions” emerge: the way frequency changes as a plane first approaches, then recedes. Two flyovers exactly 920 seconds apart.
What followed was navigation in reverse. The author took open ADS-B data: every transponder-equipped aircraft over the United States broadcasts its position, speed, and altitude at any given moment. Fifty gigabytes of tracks for a single day, and some five thousand lines of code. For every hypothetical point on the map, he calculated: which aircraft should have been audible from there at that moment, what frequency curve they would have produced, how long the sound would have taken to arrive given the terrain. The first filter left roughly 70,000 candidates. Then came an especially elegant move: if the microphone had been here, then during the “quiet” interval between the two flyovers, a third aircraft should also have been audible from that point — but it was absent from the recording. Cross that point out. Around 15,000 remained. The final comparison of simulated Doppler curves against the spectrogram produced a clear winner, and a Street View check sealed it: the shape of the tree crowns and the silhouette of a utility pole matched the frame. The camera was found.
What is especially valuable is that the author took the trouble to run a negative control on himself: he took another minute of recording with a different aircraft and ran the same search again. The independent check converged on the first answer. This was no longer a “wow effect.” It was a method.
The mechanics are trivial in exactly the way that makes them hard to defeat. Sound travels at 343 meters per second, which means you always hear an aircraft in the past: a plane at a slant range of six kilometers sounds the way it was seventeen and a half seconds ago. And in those seconds it has time to move, so a single time correction will not do — the delay has to be solved anew for every moment of the flyover. There is a subtle detail people rarely remember: altitude in ADS-B is measured above sea level, while the microphone sits on the ground, so the model needs digital terrain data to calculate the true three-dimensional distance.
The frequency the microphone hears is the aircraft’s base tone, multiplied by the harmonic number and by the factor c/(c+v), where c is the speed of sound and v is the aircraft’s velocity along the line toward you: on approach the tone rises, on departure it falls. The entire shape of that curve depends on where you are standing: the spectrogram becomes a position detector, and every point on the map has its own unique flyover “signature.”
The best part came after publication. An enthusiast at Metabunk reproduced the search independently — using the Sitrec tool, carefully described in an open guide: with no hints about the coordinates, aircraft IDs, or recording start time. Searching the full day’s archive of tracks took about ten minutes of computation and produced a point on the north shore of Long Island, New York — 38 meters from colsto’s result. The top candidate: two aircraft, a PC12 with registration N22NY (base frequency 112.2 Hz, closest point of approach — 6.1 km, speed 137 m/s) and a B350 with registration N93CQ (125.5 Hz, 3.6 km, 104 m/s). The recording’s start time was reconstructed to the millisecond: 20:17:29 UTC, September 8, 2026. Open data going in, independent replication coming out; the guide’s authors honestly note that the method’s settings were adjusted as development went along, so this was not a blind experiment — but the convergence is worth a great deal on its own.
This story has a prologue that is a pleasure to revisit. February 2014: Finnish electronics engineer Oona Räisänen, known online as windytan, sat down with a YouTube video of a police chase in Kansas City, filmed from a news helicopter. In the audio, she noticed a strange, persistent interference — and did not stop until she understood: it was not interference. Through the left audio channel, the helicopter was transmitting digital telemetry. Binary frequency-shift keying, the carrier hopping between 1200 and 2200 hertz, a stream of 1200 bits per second, 47-byte packets with 0x80 as a separator. After demodulation, the bits assembled themselves into Bell 202 ASCII text — strings like “#L N390386 W09434208.” Those were the helicopter’s coordinates: 39 degrees 03.86 minutes north, 94 degrees 34.2 minutes west — a point in Kansas City. Oona overlaid the helicopter’s reconstructed trajectory on a map and got the exact arc of its path over the chase. An update to the post put everything in place: the video downlink went through a Nucomm Skymaster-class transmitter — these can send live GPS coordinates, and they mix that channel straight into the audio.
The difference between 2014 and 2026 is telling. Back then, the “leak” had to be teased out of a protocol specification — fortunately, the telemetry turned out to be plain text. Now the leak lies deeper: the physical sound of the environment itself cuts curves across the spectrogram, and mathematics extracts a place from them. There is nothing to crack. You only have to listen.
Some Lobsters readers initially skipped the video without opening it: the title, “sound reveals location,” sounded like a retelling of an old familiar technique — forensic analysis that ties a recording to an electrical grid. That familiar technique is indeed exactly as dull as it is useful. The AC power grids of most countries do not hold their nominal 50 or 60 hertz perfectly steady — the frequency continuously drifts microscopically according to the balance of generation and consumption, and that drift is shared by every outlet in the region. The background “hum” in any recording — a digital recorder or a battery-powered phone, it does not matter — carries the imprint of the power grid’s condition at the moment of recording. Compare it with a reference database, and you get a time and a place. The technique is called ENF, electric network frequency, and it is seventeen years old: Carlos Grigoras published the method in the Journal of the Audio Engineering Society back in 2009, while the FNET monitoring network — a joint project of the University of Tennessee and Oak Ridge National Laboratory — has maintained public databases for audio forensics and edit detection for years: jumps in frequency reveal where a recording was cut.
Commenter trenchant admitted that this was precisely why he had put off opening the video — and then came away saying: “Planes are far more interesting.” That is more accurate than it sounds. ENF ties you to an electrical grid, roughly, to within a region, and it works even with a battery-powered recorder — mains hum enters the recording through induction — but it cannot do without a place with electrical infrastructure. The Doppler method requires only sky overhead and gives you a point on the map. It is also much harder to defend against: you can hide from ENF in a deep forest without power lines, but there is no way not to breathe the air in which planes are audible.
The most sobering thing about this story is how old it is. The formulas describing the frequency curve of a flying aircraft’s sound at a stationary microphone were published by Ferguson and Quinn in the Journal of the Acoustical Society of America back in 1994. In 2014, Martin and Geneske published a paper in IEEE Transactions on Aerospace and Electronic Systems: seven microphones around an airport, an ambiguity function between pairs of receivers, a genetic algorithm as solver — in simulations, positioning error came out on the order of the aircraft’s own length. In 2017, Pennsylvania State University and the Army Research Laboratory learned to track flyovers using six microphones placed too far apart for conventional beamforming, via a time-frequency matched filter — under conditions of monstrously low signal-to-noise ratio. And in 2023, AeroSonicDB appeared — an open dataset: 625 flyover recordings, 301 aircraft, labeled not by hand but passively, using the same ADS-B data. For decades, this was an engineering curiosity for aircraft-noise specialists and military acousticians. Two pieces were missing: open global tracks and cheap computing. Both arrived, and an academic curiosity became a consumer technique.
A month ago, there was a story here about Transit — the first satellite navigation system — and at its heart lay the same inversion. In 1957, Guier and Weiffenbach listened to Sputnik 1’s beeping signal and used its Doppler curve to calculate its orbit: the receiver’s position was known, the satellite was unknown. Frank McClure inverted the problem: if the orbit is known, the same curve can be used to calculate the receiver’s position. That is how navigation was born: a submarine at periscope depth listened to a flyover and learned where it was.
colsto is doing exactly the same thing, nearly seventy years later. Instead of a radio signal — acoustics. Instead of orbits transmitted from a satellite — public ADS-B tracks. Instead of a ballistic-missile submarine — a home server and a camera facing a river. The inversion is exactly the same: “I know the source — find the receiver.” Only who now plays the role of the receiver has changed.
The best comment in the thread belongs to zesterer, and it deserves to be translated in full: “Every digital artifact contains enough entropy to tell you almost everything about its creation and creator — it’s just wrapped in layers of obfuscation we can’t see. Even ‘noise’ is not noise: it is a manifestation of the physical environment. It is a signal too — depending on what you are looking for and how cleverly you know how to isolate it.”
The practical implications are drier. There is exiftool for EXIF tags; there is no equivalent “scrub it and forget it” for a Doppler trace — the leak is not inside your file but in the physical world around the recording. The author of the Sitrec guide honestly lists the method’s weak points: the model assumes a stationary microphone, subsonic speed, and an unchanging engine tone; gaps in tracks, wind, and quiet aircraft weaken the inference; the gap between the top candidate and second place is not a confidence percentage, and reliability must be judged from the curves, not the number. In colsto’s case, the gap was convincing, but that is a particular case, not a universal law. And yes — there is essentially only one way to leave no trace in sound: do not record where planes can be heard, or understand that the location is already in the recording even if it is absent from the metadata.
Which leads to the moral conclusion, unexpectedly cozy for a story about surveillance: the listening party is doing nothing forbidden. They are not hacking devices, reading someone else’s correspondence, or bypassing protections — they are listening to a world that has always spoken aloud. It is simply that now everyone has the mathematics to hear it.