Can your audio be saved?

The shoot is over. The person has gone home, the room has been put back, and the audio is wrong. The question is whether to pay somebody to rescue it or accept the loss and film it again.

A useful first assessment separates straightforward reduction, repair with audible trade-offs and material that may need recording again. These are working categories; a test on the actual recording decides the recommendation.

Most people ask what a fix costs. The more useful question is which bucket they are in, because that changes what to do this week and what to do next time.

What can be repaired at reasonable cost?

Steady noise is often a better candidate for reduction, provided enough of the wanted sound remains distinct.

That covers a mains hum, the drone of a fridge or a fan, the hiss of a recorder turned up too far, and the low rumble a building makes that you stopped hearing years ago. None of them changes while you speak, which is what makes them the tractable case: the reduction is still an estimate rather than a perfect lift, so it is never quite free, but kept gentle it costs the voice very little, and only a heavy hand starts to thin it. How clean it comes depends on how loud the noise is against the voice and how well the two are separated to begin with.

A level that is too low but otherwise clean sits here too. A quiet recording is not a damaged one; it needs bringing up, and the only question is what comes up with it.

Mild sibilance belongs here. Sibilance is the harsh spit on s and t sounds, usually produced by a microphone too close or pointed straight down the middle of a mouth. Reducing it is routine, and it only becomes irreversible when somebody overdoes it.

Repairs with a trade-off

This is where a repair needs a sample comparison. The question is whether the improvement outweighs any change to the voice or performance.

Heavy echo is the main one. Reflections from hard walls arrive after the direct sound and overlap with it, so removing them removes part of the voice they are wrapped around. It can be pulled back a long way. Push further and the voice starts to sound processed and slightly synthetic, which every listener feels and almost none can name.

An unstable delivery is the second. Somebody who drifts towards and away from the microphone, or whose energy rises and falls unpredictably, produces a recording where every sentence needs different treatment. It can be evened out, and evening it out flattens the performance, because the variation is partly the person.

A distant recording contains more reflected sound relative to the direct voice. Processing may reduce it, with limits and possible artefacts; a sample helps decide whether that trade is acceptable.

What is gone?

Two things, both common.

Clipping

Clipping is what happens when the recording level was set too high. Sound is stored as a shape, and if the shape is bigger than the space available, the top and bottom are cut flat. The parts that were cut off were never written to the file.

You can recognise it without knowing anything technical. The audio sounds acceptable at normal speech and then breaks up harshly on the loudest moments: a laugh, a clap, a pan put down on a hob. It is a crackle that appears only at the peaks. If you look at the waveform, the jagged shape your editing software draws to represent the sound, the loud sections have flat square tops where there should be points.

Software can smooth the edges and make it less unpleasant to sit through. Newer repair tools go further and rebuild a plausible version of the flattened peaks by inferring what belonged there, which is worth trying on a moment you cannot reshoot. It is reconstruction rather than recovery: the file never held that shape, so the tool is proposing one, and on anything that matters the fix is still to record it again. That line, between repairing what exists and inventing what does not, is the exact spot where money stops working.

When the room is louder than the person

Removal leans on there being enough clean, direct sound to hold on to. When the traffic, the air conditioning or the echo competes with the voice at similar strength, there is little clean signal underneath to protect: a plain noise reducer takes the voice out with the noise, and even the newer separators that prise a voice from a room hand back something thin and uncertain. Sometimes that is enough; more often, when the room is that loud, filming it again is cheaper and comes back better. Leave the noise in and the words will not survive a phone held at arm's length on a bus.

A distant phone recording in a reflective room is worth testing before paying for a full repair. Where the result remains unclear or unnatural, a retake may be the better use of the budget.

How do you tell which bucket you are in?

Listen to the gaps.

Play the recording and pay attention to the silences between sentences rather than to the speech. Whatever sits in those gaps is your noise, and how loud it is against the speaking tells you almost everything. Quiet gaps put you in bucket one. Busy gaps with the words still cutting through put you in bucket two. Gaps nearly as loud as the words put you in bucket three.

Check loud moments on headphones and another device. Crackle can indicate clipping or another recording or playback fault; confirm the source before deciding the remedy.

Why the next shoot is the cheap place to fix this

Add up what is in the three buckets. A great deal of what goes wrong with audio was decided before anybody pressed record: how far away the microphone was, what level it was set to, whether the fan was running, whether the room had anything soft in it.

Before the next shoot, test microphone position and recording level, listen for appliances and make a short sample. Leave headroom for louder moments. Existing equipment may be enough to make a substantial improvement.

So the question after a bad recording is whether the thing that put it there will be in the room again next time.

Sources and evidence

What here is established signal processing and what is our own method.

What is established. What can and cannot be recovered follows from how the fault sits in the signal: steady, separable noise can be reduced, while a problem that overlaps the voice is estimated rather than restored, and reducing it trades against the speech itself (peer-reviewed speech-enhancement research, Loizou and colleagues). Modern tools push that trade further; they do not remove it.

Our method. Judging each rescue by whether the voice still sounds like the person, and saying when re-recording is the cheaper answer, is the practice here.

References.

• Kim, G. and Loizou, P. C. Gain-induced speech distortions and the absence of intelligibility benefit with existing noise-reduction algorithms. pmc.ncbi.nlm.nih.gov . Noise suppression introduces amplification/attenuation distortion; conventional NR does not by itself raise intelligibility. Constraining amplification distortion can yield gains; does not establish a viewer-retention effect.