The noise worth leaving in

A synthetic comparison makes the cost of indiscriminate cleaning audible: it can also remove the detail we wanted to preserve.

Cinema and audiovisual media
A sieve drops fragments and catches a thin blue thread.
A sieve drops fragments and catches a thin blue thread. AI-generated illustration: a conceptual scene, distinct from the example on the page.

The background drops and the sound immediately seems tidier. The satisfaction is quick: something disturbing has been removed. But if we listen only to the silence between two events, we may not notice what we lost within the weaker event.

Cleaning a recording requires deciding what counts as noise. The choice also depends on its use. A rustle may obscure a word, or preserve the continuity of an environment that the edit needs us to hear. The same threshold does not automatically serve both purposes.

An example that discloses its construction

The example contains a synthetic signal with a background and two impulses of different intensities. In the second version I apply an amplitude threshold: samples below it are set to zero. This is a deliberately elementary treatment, rather than a simulation of professional software’s algorithm.

Before and after the threshold

Original synthesised audio, without speech.

Original synthetic signal · 4 s
Amplitude threshold applied · 4 s

The comparison makes the problem of selection audible. The strong signal retains a recognisable part, while the weak one may be cut away with the background. There are no recorded voices, places or testimonies; both files are generated for this article and have the same four-second duration.

The amount of noise removed is not enough to judge what remains to be heard. We need to return to the event we meant to preserve, especially in less obvious passages. A perfectly still background can accompany an already damaged sound.

What the threshold does not know

The Audacity manual describes limits to noise reduction when useful sound and noise resemble each other or are difficult to separate. My demonstration threshold does not use its algorithm: it merely makes a selection problem coarse and visible.

Two folded strips compare a textured surface with a smoothed one.
Two folded strips compare a textured surface with a smoothed one. AI-generated illustration: a conceptual scene, distinct from the example on the page.

A threshold looks at a value, rather than the meaning of what produced it. Even a more sophisticated procedure must rely on signal characteristics. Judging the result therefore needs something a cleanliness indicator alone cannot decide: which detail matters and how difficult it has become to recognise.

A comparison can also mislead if one version is simply louder. This example applies no separate normalisation to bring each file back to the same peak. The transformation is disclosed and the examples should be compared without deliberately changing volume between them. It is not a controlled perceptual test.

Keeping a way back

The AI-generated illustrations show a sieve and a smoothed surface. They are not waveforms or spectral analyses. The sound and exact transformation remain in the files and generation code.

When working on a recording I would keep the original and a note of the processing. A pause heard again after editing might require a different background, or restore value to a detail discarded too soon. Without the source material, that correction becomes a search in the dark.

A good intervention preserves enough of a sound to sustain the reason we recorded it. The noise worth keeping is no romantic virtue of dirt: it is the part of the material that impatient cleaning would cease to distinguish from disturbance.

Two folded strips compare a textured surface with a smoothed one.
Two folded strips compare a textured surface with a smoothed one. AI-generated illustration: a conceptual scene, distinct from the example on the page.