Environmental acoustics · auditory scene analysis

When the Room Becomes the Sound

Seven people carried a phone for two weeks. Twenty-four times a day it asked them where they were, and afterwards someone went to each place with a speaker and a recorder and measured it. The result is 270 impulse responses of ordinary life: bedrooms, a car, six bathrooms, the stairwell of an elementary school. Every one of them blurs sound in almost the same way, and your hearing quietly undoes it. Below, those rooms are measured again from the published audio, and then rebuilt in your browser out of nothing but the measurements, so you can hear both the undoing and what happens when it fails.

Put a sound in a room

Nothing recorded in these rooms is played here. Each room arrives as sixty-six numbers, and is rebuilt from them.

 

The sound

as measured

Ready. Headphones help.

Reverberation time by frequency. Faint lines: every surveyed space. Bright line: the room selected above, as measured from the published recording. Press measure what you just heard to add, in yellow, the profile recovered from the audio this page synthesised. With the slider away from as measured the yellow line is the room you asked for and not the room that was recorded, so the two are supposed to differ.

Sixty-six numbers is a room

The room you just heard was not a recording. For each of 33 frequency bands the page holds two numbers, measured from the survey's own audio: how long that band takes to fall by 60 decibels, and how much energy it carried. To rebuild the room it fills those bands with Gaussian noise, multiplies each by a decaying exponential of the measured length, filters them again, adds them up, and puts a single impulse at the front for the sound that reaches you straight from the source.

That recipe is not a convenient approximation. It is what Traer and McDermott did, and they tested it on people: twenty-two listeners heard a real recorded room and a room synthesised this way with its decay and direct-to-reverberant profiles matched, and were asked which was real. They could not tell, for any of three kinds of source (t(21) = 1.52, P = 0.14 for impulses, -0.45, P = 0.66 for speech, -0.87, P = 0.40 for noise). Invert the frequency dependence instead and they caught it every time. That is a null and an effect side by side, which is a stronger thing than either alone: what a space does to a sound, once the first few reflections are past, is in how fast each frequency dies, and in nothing else the ear can hear.

The slider is the experiment

Every real room in the survey rings longest somewhere in the middle of the spectrum and dies fastest at the top: air and soft furnishings eat treble. That is the regularity the slider destroys. At as measured each band decays at its measured rate. At none every band decays at the same rate, which is a room with no colour. Pushed to inverted, the treble rings longest and the middle dies first.

 

The survey's authors built exactly this manipulation, put it in front of listeners, and wrote down what happened:

IRs with unnatural frequency dependence (i.e., spectrally inverted) often seemed to contain two sounds: a source with moderate reverberation and a high-frequency “hiss.” The auditory system is apparently unwilling to interpret high frequencies that decay more slowly than low frequencies as reverberation, ascribing them to an additional noise-like sound source rather than an impulse interacting with the environment.

That is the whole point, and it is why hearing works at all. A room and a sound arrive at your ear added together, and pulling them apart is an underdetermined problem: any sound in any room could be some other sound in some other room. It is solvable only if you already know something about rooms. You do. And when a room violates what you know, the part that does not fit stops being a place and becomes a thing making a noise.

What happens for you at the far left of that slider is a demonstration, not evidence. The evidence is theirs, with twenty-two listeners in a soundproof booth. This page can only hand you the stimulus.

Measuring 270 rooms again

None of the numbers above are taken on trust. The survey's audio is committed to this project's repository, unmodified, and every figure on this page is measured from it by code you can read, which is the same code your browser just ran to rebuild the room. The measurement is a Schroeder backward integration per band, fitted between −5 and −35 decibels, anchored where the response stops being a handful of echoes and becomes diffuse noise.

 

 

How much of a room is the same as every other room?

 

The reverberation of an ordinary hour

The survey has something in it that its own paper does not use. Because the spaces were found by texting people at random moments and asking where they were, each recording carries a count: how many of those moments happened there. That count is a sample of how a life is distributed across rooms. Weighting by it turns a survey of rooms that exist into a measurement of the acoustic a person is actually in, which as far as we can find nobody has computed.

 

Where the sampled moments of seven lives actually happened, by how long the room rings. The bar under a third of a second is mostly somebody's bedroom.
Every room type in the survey with at least three measurable spaces, ordered by how much of the sampled fortnight happened in it.
Kind of spaceSpacesSampled momentsMedian RT60Range

The survey's five findings, checked

Traer and McDermott state five regularities. Each one is a claim that can be recomputed from the audio they released. Four reproduce. One is confirmed in the half of it we can test and contradicted in the other half, and the reason is in the right-hand column.

Their claimVerdictWhat we measured

Against their own published numbers

 

What this page refuses to tell you

 

A reverberation time is only measurable when the recording contains one. Outdoors there is almost no diffuse field, so what follows the direct sound is the measurement's own noise floor, and fitting a decay to that measures the equipment. A few of the released files are two tenths of a second long in rooms that ring for most of a second, so the decay simply is not in the file. Both cases are refused rather than estimated, and the count of refusals is above. An earlier version of this analysis did estimate them, and disagreed with the survey by a factor of twelve on a street corner and a factor of ten in an atrium, in opposite directions.

Three counts of one survey

 

The check

 

The strongest check is one you can run: press measure what you just heard. The page takes the audio it synthesised, runs the measuring code over it as if it were a recording of a real place, and draws the recovered profile against the one it was built from. If the reconstruction were decorative rather than real, those two curves would not land on each other.

Everything here is reproduced by verify-when-the-room-becomes-the-sound.mjs in the repository, offline, from the committed audio.