🔊 Planes Leak Your Location: How Inaudible Doppler Curves Geolocate a Recording

What the video shows

The challenge is a one-minute clip from a security camera. The picture is a blurry tree line. The only facts given are the date, 8 September, and that it is somewhere in the continental United States colsto, “I discovered a sound that leaks your exact location” (YouTube, 2 Nov 2026). He finds the camera.

The clue is inaudible. On a spectrogram, zoomed into low frequencies at high resolution, two faint curves appear. Each is the engine tone of an aircraft sliding down in pitch as it passes: a Doppler curve. The two flyovers are 920 seconds apart colsto, “I discovered a sound that leaks your exact location” (YouTube, 2 Nov 2026).

The method, step by step

1. Get every flight. He downloaded the ADS-B tracks of every aircraft over the US that day, several hours of scraping and files in the tens of gigabytes colsto, “I discovered a sound that leaks your exact location” (YouTube, 2 Nov 2026).
2. Keep the propeller planes. The tone is a propeller signature, so jet types were filtered out using an aircraft-type lookup table.
3. Lay a grid. Points every 200 metres across the search area. For each point and each flight, compute the moment the plane was closest, the closest point of approach (CPA). Then add the time the sound takes to arrive over the slant distance. Altitude is above sea level, so a digital elevation model supplies the ground height at every point.
4. Read speed and distance off the curve. How far the pitch drops gives the plane’s speed. How fast it drops gives the CPA distance. Both become filters.
5. Filter by timing. Keep only points that would hear a plane of the right speed and distance, then another one 920 seconds later. Then drop points that would have heard a third plane in between. About 70,000 candidates fell to 15,000 colsto, “I discovered a sound that leaks your exact location” (YouTube, 2 Nov 2026).
6. Model and score. For each surviving point and flight, compute the exact Doppler curve that point would have heard. It comes from the plane’s real position and radial velocity, second by second. Draw it on the spectrogram and sum the brightness under it. Refine to a 100-metre grid.
7. Confirm on the ground. One flight pair scored far above the rest. Street View from that spot showed the tree line, and then the camera itself colsto, “I discovered a sound that leaks your exact location” (YouTube, 2 Nov 2026).

A second test used one flight and a tight time window instead of two flights. It found the same spot. Either a second flyover or a known recording time is enough; you need one of them, not both colsto, “I discovered a sound that leaks your exact location” (YouTube, 2 Nov 2026).

He describes it as over 5,000 lines of code, built over months, and says: “No code or tools are being released” colsto, “I discovered a sound that leaks your exact location” (YouTube, 2 Nov 2026).

Why it works

• Low frequencies travel. Engine tones sit at roughly 50–300 Hz. They pass through walls and are recorded by every phone, camera and doorbell, whether or not anyone hears them.
• Doppler is geometry. The shape of the pitch curve depends only on the plane’s speed and how close it passed. A curve is a measurement of distance.
• The plane’s side of the equation is public. ADS-B broadcasts position, altitude and speed in the clear. Hobbyist networks archive them.
• Two curves beat one. The time between two flyovers, plus the silence between them, rules out most of the country before any heavy computing starts.

The physics is old. Locating aircraft from the Doppler shift of their sound is published work. A 2014 IEEE paper does it with seven fixed microphones around an airport Martín Román et al., IEEE Trans. Aerospace and Electronic Systems (2014). A 2026 review covers Doppler localisation across radio, acoustic and satellite settings Szczepanik, arXiv:2604.14413 (v2, 7 Jun 2026). What is new is turning it around: the aircraft positions are known, so the microphone is the unknown.

Could this be automated? An honest assessment

We were asked whether Max Intel could build a browser tool for this. We did for ADS-B decoding and weather chronolocation. Here is the breakdown.

Our answer: a full “upload audio, get coordinates” tool is not something we will build, and not only for compute reasons. The output is a private location from a recording the subject may not know exists. That is the definition of the harm our ethics guide draws a line under. The author reached the same place: he kept the code to himself.

We built the defensive half instead. Scout Doppler scans your own recording for Doppler curves and tells you whether a flyover is in it before you post. It also runs live from a microphone as a flyover logger, matching each pass to ADS‑B when you have a receiver. It estimates speed and closest distance only. It never searches for where a recording was made.

For investigators with a lawful purpose and a small search area, the recipe above is enough. It can be reproduced with Python, the adsb.lol archive and a DEM. The ADS-B data has the same fields our radar page decodes live.

Where it fails

• No planes, no fix. Remote areas and quiet hours give nothing to match. Jets at cruise altitude leave weaker, flatter curves than low propeller traffic.
• Coverage gaps. ADS-B archives depend on volunteer receivers. Aircraft below receiver horizon, or in the FAA’s privacy programmes, may be missing or anonymised.
• Search area times time window. With no idea of place or time, the search becomes a nationwide, all-day grid. The video needed two flyovers and a known date to make that tractable.
• Edited audio. Cuts break the timing between events. Heavy processing can smear the curve.
• It finds the microphone, not the person. A security camera, a phone on a desk, a dashcam: the fix is where the recorder was.

What actually protects you

He tested the obvious defences on camera, and most fail colsto, “I discovered a sound that leaks your exact location” (YouTube, 2 Nov 2026).

• A high-pass filter does not work. The curves share frequencies with the human voice, so cutting them cuts speech. And a cut curve came back when the quiet band was boosted.
• Talking over it does not work. A speaker playing a Doppler curve across the room still showed on the spectrogram under his voice.
• Phone calls do not strip it. The curve survived a phone’s compression and noise reduction. A call is also a recording with a known time, which is the one thing that makes a one-plane search easy.
• Cuts work. Removing sections breaks the time offset between flyovers. Posting a trimmed clip is safer than posting a continuous one.
• Noise cancellation helps. Aggressive noise reduction aimed at steady tones can remove the curve. Check the result on a spectrogram, not by ear.
• Delay and vagueness help. Not revealing the recording date, or the rough region, multiplies the search cost. This is the same advice as our travel guide gives for photos.

The bigger picture

This sits in a line of work that reads location from the environment rather than from metadata. Weather and shadows give the time of day; our chronolocation tool uses the first. Mains hum, the 50 or 60 Hz electrical network frequency, has been used in forensics for years to date recordings. Thunder gave colsto about five kilometres in an earlier video; aircraft give him a hundred metres colsto, “I discovered a sound that leaks your exact location” (YouTube, 2 Nov 2026).

Each of these needs a public reference dataset. Flight tracking is the richest one there is: free, archived daily, and global. That is why the ADS-B guide and the Remote ID guide matter to privacy as much as to plane-spotting. The data that lets you watch a plane is the data that lets a plane watch you.