📍 The claim 0:00
The challenge: a static video with no visual change, no landmarks, no metadata worth the name. The only given information is a date (8 September) and "somewhere in the continental US". The goal is a street-level match.
The tool is not a landmark or a shadow. It is the Doppler curve of aircraft flyovers, present in the audio at frequencies you cannot hear, at a level you cannot notice.
🔊 Why Doppler beats thunder 0:41
The same researcher previously geolocated recordings by thunder. Thunder has excellent reach — it travels far and penetrates structures — but the best it gave was around 5 km of precision: enough for a city and a side of town.
| Signal | Reach | Precision | Why |
|---|---|---|---|
| Thunder | Very good | ~5 km | A single boom with no fine structure to fit |
| Aircraft Doppler curve | Very good | Tens of metres | Shape encodes speed and closest-approach distance; timing encodes geometry |
The precision comes from the fact that both benefits survive: low-frequency sound travels and penetrates walls, and the curve's shape is a two-parameter physical measurement rather than a single event.
🌈 The spectrum nobody looks at 1:06
Two clips are played: one contains the geolocating information, one does not. At normal playback they are indistinguishable, because the signal is inaudible. In a spectrogram — time on X, frequency on Y — it is still invisible at default zoom and resolution. Only when you zoom into the right band and raise the resolution does a thin descending-then-rising trace appear: a Doppler curve from a propeller aircraft.
✈️ Step 1: narrow by flight 2:37
The prior art that makes this tractable is ADS-B: publicly broadcast aircraft position, altitude, velocity and identifiers, available through commercial APIs and through scraping. The author pulled every flight over the continental US on the target date — tens of gigabytes of parquet — and immediately cut it down.
- Propeller, not jet. Listen to the boosted clip and the aircraft type is obvious to the ear. ADS-B's type code field does not say "prop vs jet" directly, but ICAO type designators do — and since that lookup table is incomplete, he inverted the logic: filter out all known jet types rather than trying to whitelist props.
- Everything else is a candidate. The resulting map is still the whole country with a lot of lines on it.
📐 Step 2: CPA and the delayed CPA 4:31
Terminology, first — and one correction: the video calls CPA the "closest point of arrival". The standard term is closest point of approach.
CPA matters because "hearing a plane" is not an instant. A flyover's Doppler curve lasts close to two minutes, so the usable timestamp is the CPA time — the moment the aircraft was closest, when the curve crosses its fundamental and is loudest.
Then two physical corrections turn a rough idea into arithmetic:
| Correction | Why |
|---|---|
| Delayed CPA | Sound takes time to travel. The listener hears the CPA later than it happened — seconds later at kilometres of range. For a point 3+ km away, the example goes from a CPA of 22:46:36 to a delayed CPA of 22:46:46. |
| Slant distance | Aircraft altitude is above sea level, not above the ground. To get the true straight-line distance you need terrain elevation at each candidate point — which means a digital elevation model (DEM), not just a flat map. |
Now lay a 200 m grid over the search area — he uses a generous 10 km radius around each flight, because beyond that you would not hear an aircraft even in ideal conditions — and for every grid point record the delayed CPA time for that flight. Do it for all flights and each point holds a list: at these times, you would have heard a plane.
⏱ Step 3: the 920-second offset 7:04
The recording contains two flyovers. Take the offset between their CPA times — about 920 seconds — and keep only the points that heard a plane and then, 920 seconds later, heard another.
Note the honest treatment of error: where exactly you measure the offset from varies by a few seconds, so the filter allows a window rather than an exact match. That kind of small, declared tolerance is the difference between a filter that works and one that eliminates the true answer.
| Filter | Candidates | Note |
|---|---|---|
| Time offset only, small test area | ~35,000 | That area is under 0.05% of the continental US |
| + velocity and CPA-distance bounds | ~73,000 | Clustered around airports — which is a clue in itself |
| + silence filter | ~15,000 | Still far too many to model |
🎚 Step 4: the curve is a fingerprint 8:28
Two clips, one audibly faster. That intuition is the whole physics: a constant-velocity flyover produces a Doppler curve with exactly two parameters.
| Parameter | What it controls in the curve |
|---|---|
| Aircraft velocity | The total frequency drop — how large the excursion is |
| CPA distance | How quickly the drop happens — the sharpness of the transition |
So instead of filtering for "heard a plane, then 920 s later heard another", you filter for "heard a plane at this speed from this distance, then 920 s later heard another at that speed from that distance". Those two parameters collapse a hundred thousand candidates into tens of thousands.
Sanity check on the numbers: for a source at 130 Hz moving at 100 m/s, the Doppler formula f = f₀·c/(c + v_r) predicts the received frequency falling to about 101 Hz as it recedes — a drop of roughly 30 Hz over the flyover, which is exactly the scale of the traces in the spectrogram.
🔇 Absence is also signal 11:36
Most of the remaining candidates cluster around airports — because that is where aircraft are. But the recording contains almost 15 minutes of no Doppler curve between the two events, and you do not get silence like that next to a major airport.
🎯 Step 5: the mistake that breaks everything 14:15
Fifteen thousand candidates still meet every constraint, and the spectrogram has been squeezed dry. The next step is deeper: instead of reconstructing a flight path from a curve, model the exact curve each candidate flight geometry would have produced, then score observed against modelled.
Modelling is straightforward in principle. The Doppler curve is a function of the aircraft's radial velocity — how fast it moves directly toward or away from you. Head on, radial equals true velocity; directly overhead, radial is zero and there is no shift. Sample the flight data at successive times, compute radial velocity, push it through the Doppler equation, plot the resulting frequency.
And here is the trap, which he pauses on deliberately:
Naively, you would compute frequency from the aircraft's state at emission time minus nothing, or you would compute the correct curve and then delay it. Both are wrong in the same way, and he demonstrates why: shifting the correct curve to line up one point makes the rest diverge badly. It is not the same curve, late — it is a different curve.
🧮 Step 6: locked scoring, 100 m grid 18:18
Scoring is simple to state: overlay the modelled curve on the observed spectrogram and sum the brightness it covers. Two sources of error need handling, because both the fundamental frequency and the CPA time were measured by eye from the curve.
| Event | How it is scored |
|---|---|
| Event 1 | Place the modelled curve at the measured CPA timestamp, then sweep a few seconds of time offset and a range of fundamentals, and take the best combination |
| Event 2 | Sweep the fundamental only. Its temporal placement is fixed by flight geometry, so if it lands at the wrong offset after event 1 is aligned, that is genuine evidence against the candidate |
That asymmetry is the clever part: correct the measurement error once, then let the physics carry the second event. The 920-second offset becomes a hard physical constraint rather than a free parameter that could be tuned into existence.
🎬 One minute, unknown place 22:47
The first search resolved by ranking: score each candidate on both events, group by flight pair, keep the best point per pair. One point stands out at 20.48, with the runner-up around 15 — and the gap is visible in the spectrograms, not just the numbers. First place covers the observed curve almost exactly; second place misses the top of event one and is plainly wrong on event two.
Then the harder demonstration, and the one that should worry anyone who publishes raw audio:
| Given | Constraint | Result |
|---|---|---|
| Two flyovers, unknown place, unknown time | 920 s offset | One clear winner |
| One single minute of audio, known timestamp, US-wide | None on location | Same location, first place by a wide margin |
Two entirely different methodologies — two events and no clock, versus one event and a clock — put the same point first. The trade-off he draws out is the useful rule: you need one of the two constraints to be good. A city is about 0.01% of the land area, and constraining space that tightly lets you relax time completely; you may not need the minute, the hour, or even the day.
🏞 The reveal 25:33
Back to the map: the winning coordinate lands on a shoreline. With flight IDs known, he re-scores at much finer resolution around each flight path, producing heat maps per event whose intersection is a small patch of bank. Then street view: find the V-shaped tree, the point beside it, the rounded mass after it — and, at the top of a pole, the security camera that recorded the clip.
🛡 Defences, and why the obvious ones fail 26:09
The last third is the part to actually act on, and it is a sequence of intuitive defences being demolished on camera.
| Defence | Result | Why |
|---|---|---|
| High-pass filter to cut the low frequencies | Fails twice | The curves sit in roughly the same band as the human voice — you would remove what you want to keep. And after exporting to MP3 and boosting the remaining bands, the curve returns |
| Drown it out with foreground noise | Fails | A speaker across the room playing a Doppler curve is still measurable under speech — he demonstrates this live |
| Phone compression and noise reduction | Fails | Routing the curve through a real phone call — with the strong signal muted so only the phone's output is audible — the curve survives and even leaves a visible gap when it stops |
| Cutting the audio | Partially works | Removing the segment destroys the two-event offset. But if you keep a single minute and the location search space is large, you then need a precise recording time |
| Additional noise cancellation | Practical, untested | His suggestion, not a demonstrated result |
⚖️ The honest threat model 28:09
He explicitly refuses to fearmonger, and the caveats are the most important lines in the video:
- The odds are slim. Few people can do this and fewer will.
- There are a hundred easier ways to find you than fitting Doppler curves to aircraft telemetry.
- The entry cost is real: months of work, thousands of lines of code, tens of gigabytes of parquet, 3D physics, country-scale elevation models.
- But the plausible use case is not casual. A recording with literally zero visual clues — footage from an investigation, a human video online, a random security camera — becomes locatable.
🧠 What generalises beyond this 29:05
Strip away the aircraft and this is a template for any physical-inference search:
- Fuse a public dataset with a raw sensor side channel. ADS-B is public; the audio is yours; the intersection is a location.
- Filter in order of cheapness. Type, then time offset, then shape, then full modelling — each stage exists so the next one is affordable.
- Treat absence as evidence. Silence between events eliminated more candidates than the events themselves.
- Model forward, don't invert backward. Reconstructing geometry from a curve is ambiguous; modelling a curve from known geometry is not.
- Lock what physics fixes, sweep what you measured. Correct the measurement error once, then hold the derived constraint rigid.
- Respect propagation delay. The state you observe now is the state at emission.
- Publish the method, withhold the tooling. A rare and, here, defensible choice: the awareness travels, the capability does not.
✅ Claims checked
| Claim | Verdict | Evidence |
|---|---|---|
| ADS-B provides public aircraft position, altitude and velocity | Correct | Public broadcast; commercial APIs and scraping both viable, as described |
| Doppler curve shape depends only on velocity and CPA distance (constant velocity) | Correct | Standard result for a moving source: f = f₀·c/(c + v_r) |
| Slant distance requires terrain elevation, i.e. a DEM | Correct | Aircraft altitude is above sea level, not above ground |
| CPA stands for "closest point of arrival" | Correction | Standard term is closest point of approach |
| Thunder geolocation gave ~5 km precision | His own earlier result | Consistent with his previous work; not independently verified here |
| Long-wave audio penetrates walls and survives codecs | Demonstrated in-video | High-pass removal, loud foreground, and a live phone call all failed to destroy the curve |
| Filter counts: ~35,000 → ~73,000 → ~15,000 → 30,000–40,000 after refinement | His reported figures | Stated on screen per stage; not reproducible without his tooling, which is not released |
| One minute of audio located the site within a couple hundred feet | His reported result | Same location ranked first under two independent methodologies |
| Football-field arithmetic: continental US end-to-end wraps the Earth and reaches the Moon and back ~200 times | Checks out | ≈1.51 × 10⁹ fields × 109.7 m ≈ 1.66 × 10⁸ km ≈ 4,100 Earth circumferences ≈ 215 Moon round trips |
| Phone compression "surely" destroys such signals | Disproved on camera | The curve survives a real call and pauses where the source was muted |
| High-pass filtering removes the information | Disproved on camera | Export to MP3 and re-boost restores the curve |
| No code or tools released | Stated by the author | Video description: "No code or tools are being released" |
Basis note. Built from the full 29:12 transcript. The physics, the terminology and the filter chain are checked against standard results; the pipeline's performance is the author's own reported result and is not independently reproducible, since the tooling is withheld. Two of the video's own intuitions (high-pass filters, phone compression) are disproved by its own demonstrations. The located coordinates are intentionally omitted.
🔑 Key takeaways
- The audio contains more than the sound. A spectrogram of an ordinary recording can hold a location fingerprint that no one in the room could hear.
- Two parameters, one fingerprint. For a constant-velocity flyover, velocity sets the size of the frequency drop and CPA distance sets how fast it happens — enough to locate a listener to tens of metres.
- Public dataset + private sensor = location. ADS-B telemetry is free; the recording is the only piece the attacker needs.
- Silence is signal. Fifteen minutes without a flyover eliminated candidates that every positive feature had kept alive.
- Invert the incomplete list. When a lookup table is missing entries, filtering out the known-bad beats whitelisting the known-good.
- Propagation delay is not a shift. The frequency you hear now belongs to the source's state when it emitted — model forward from emission time or the whole search silently fails.
- Lock what physics fixes; sweep what you measured. Correct measurement error once, then let the derived constraint between two events stay rigid.
- Negative evidence deserves the same standing as positive evidence — and it is usually the thing nobody encodes.
- High-pass filters do not save you; the curves sit in the vocal band, and re-boosting the surviving spectrum brings them back.
- You cannot drown it out — a quiet, structurally coherent signal survives loud foreground audio and perceptual codecs.
- Codecs remove what ears hear, not what instruments measure. Masking is a property of perception; the attack is a property of spectra.
- Cutting audio helps only if you also stop needing the clock. One surviving minute plus a precise timestamp is a full search.
- Constrain one axis well and the other can be free: know the city to 0.01% of the land area and you may not need the hour, the day, or the minute.
- Publish the method, withhold the tooling. Awareness scales; capability should not.
🔗 Resources
- The source video — colsto, "I discovered a sound that leaks your exact location (and it's everywhere)", 29:12, 2 November 2026
- The full recording — the challenge footage, linked by the author
- Background reading: ADS-B as a public data source, ICAO aircraft type designators, digital elevation models, and the Doppler equation for a moving source