→ all versions of this pipeline
A phone resting against a shirt makes a sound. Eight years of bedtime stories were recorded that way, and nothing in the audit could see it. This is what it turns out to be when you measure it — and, just as usefully, what it is not.
Why this needed measuring at all. The listener could hear the noise instantly and name it. Every automatic check reported the recording clean. That gap is not a small calibration problem — it means the instruments were looking for the wrong shape entirely, and no amount of tuning fixes a detector aimed at the wrong thing.
So the sound was characterised the only way that carries any weight: against 19 events a human confirmed are the noise, and 13 spans the same human confirmed are nothing. Without that second set, every number below would be unfalsifiable.
An unpitched, sustained, broadband texture. Not a click, not a thump, not a knock — closer to a hiss with body underneath it. It fades up and it fades down.
The strongest frequency is worth pausing on. Published measurements of rubbing-type handling noise put it at 299 ± 95 Hz. This material measures 305 Hz, on a completely different microphone, in a different decade. That is about as clean a corroboration as this kind of work offers: what the listener has been hearing is textbook handling noise.
Energy across the whole spectrum, measured in decibels above each recording's own quiet background. The pale bars are the sounds the listener said were nothing — and the gap between the two is the entire detection problem.
| band | 40–120 | 120–300 | 300–800 | 800–2k | 2–5k | 5–9k |
|---|---|---|---|---|---|---|
| cloth ruffle | 14.3 | 15.2 | 22.2 | 21.5 | 19.1 | 18.7 |
| “nothing there” | 6.5 | 17.5 | 17.5 | 12.1 | 6.6 | 8.6 |
| difference | +7.8 | −2.2 | +4.7 | +9.4 | +12.5 | +10.1 |
It is wider, not louder. The sounds a listener dismisses are mid-band events — they live between 120 and 800 Hz and die away above that. The ruffle keeps going: nine to twelve decibels more energy above 800 Hz, and nearly eight decibels more below 120 Hz. It occupies both ends of the spectrum that ordinary background noise leaves empty.
That is the discriminator. Not level, not sharpness — width.
The more useful half. These are metrics that fail to separate the noise from the sounds a listener calls nothing, scored 0 to 1, where anything under about 0.3 is noise in the statistical sense rather than signal.
| metric | cloth ruffle | “nothing” | separation |
|---|---|---|---|
| peak level above the room | 39.0 dB | 38.0 dB | 0.17 |
| kurtosis (how impulsive) | 7.0 | 8.2 | 0.11 |
| onset sharpness | +0.5 dB | −0.2 dB | 0.13 |
| decay rate | 0.7 dB/s | −3.6 dB/s | 0.11 |
| loudness above the room | 23.2 dB | 16.7 dB | 0.85 |
| spectral flux | 630 | 550 | 0.85 |
It has no attack and no tail. Kurtosis, onset sharpness and decay rate all score between 0.11 and 0.13 — statistically indistinguishable from background. This single fact explains a long run of failed attempts: declickers, de-plosive tools, knock detectors and every transient-shaped repair are built to find a sound that starts suddenly and rings away. This one does neither. It is a texture, and it was being hunted as an event.
And peak level tells you nothing. At 0.17 separation, how loud the loudest moment gets is very nearly useless for identifying it. What matters is the average lift and how wide it is — 23 dB above the room across the full spectrum, sustained for a few hundred milliseconds.
The academic description of rubbing handling noise says it holds “a consistent spectral distribution over time” — a steady texture that keeps its shape. Measured here, the opposite is true: the ruffle has higher spectral flux than the background it gets confused with. Its shape moves more, not less.
Two readings, and it is not yet settled which is right. A phone pressed against a shirt may genuinely behave differently from a lapel microphone clipped to a collar. Or the published figure comes from isolated noises recorded in quiet, and does not survive contact with a recording that also has a father and two children talking in it.
A defect class with no instrument does not read as a problem. It reads as a clean recording. The audit for this material reported a single finding on a capture that a listener could hear at least three faults in — and nothing was wrong with the checks that ran. The class simply had no detector at all, so its score was a statement about the instruments rather than about the audio.
The practical consequence is narrow and useful: stop looking for a transient. Look for a wide, sustained, unpitched lift — and look for it in the pauses, because every confirmed event in this material sits between sentences with no voice inside it at all.