What can be repaired, and what cannot

The triage list

Four categories: fully repairable, partly repairable, repairable at a cost, and gone. Knowing which is which saves hours and saves promising a client something you cannot deliver.

Fully repairable

Mains hum. At exactly 50 or 60 Hz plus harmonics. A notch filter or a de-hum module removes it almost completely, because it is narrow, constant and at a known frequency.

Level problems. Too quiet, uneven, one channel only. Clip gain and channel mapping, covered earlier.

Clicks and pops. Individual sample-level glitches. Any click removal tool interpolates across them convincingly, because you are replacing a few milliseconds using the material either side.

Plosives. The thump when a p or b hits the microphone. A steep high-pass filter applied only to that word — 150 Hz for a few hundred milliseconds — removes the thump and leaves the word. Do it as a clip-level EQ on a short split rather than across the whole track.

DC offset. The whole waveform sitting above or below the centre line. A one-click fix in Audacity and most editors, and worth doing because it costs you headroom.

Partly repairable

Steady background noise. Hiss, air conditioning, a computer fan. Spectral subtraction, covered in the earlier lesson: sample the noise alone, build a profile, subtract. Six to eight decibels of reduction is usually safe. Beyond that you get musical noise — the watery, fluttering artefact where quiet harmonics of the voice fall below the estimate and are removed inconsistently.

The rule stands: if you can hear the noise reduction, there is too much of it.

Reverb. Partly, with the caveats from the previous lesson.

Mild clipping. Declipping tools interpolate the flattened tops of the waveform using the shape either side. On brief, mild clipping the result can be genuinely good. On sustained clipping the tool is inventing most of the waveform.

Wind. The low rumble comes out with a high-pass. The roaring, saturated part where the wind overloaded the capsule does not, because the microphone stopped recording the voice during it.

Repairable at a cost

Two microphones with phase problems. When two microphones record the same source from different distances, combining them produces comb filtering — a hollow, phasey sound caused by frequencies cancelling where the two arrivals are out of step.

Three fixes, in order: use one microphone only, which is usually the right answer; nudge one track by the delay between them, which you can measure from the waveform; or invert the polarity of one and see whether it improves, which it will if they were wired oppositely.

Sibilance. De-essing, covered earlier. It works, and overdone it produces a lisp.

Traffic and intermittent noise. A passing motorbike shares frequencies with the voice and is not steady, so spectral subtraction cannot profile it. Options: cut around it and use a different take, cover it with music or ambience, or accept it. Some spectral repair tools let you paint out an isolated event on a spectrogram, which works well for a discrete sound over a gap and badly for one over speech.

Gone

Say these to a client on day one:

Heavy clipping. The peaks were flattened at recording. There is no information in the flat part to recover, only plausible guesses.

Speech below the noise floor. If the voice is quieter than the hiss, no amount of processing separates them, because there is less voice than noise in every band.

Two people talking over each other on one microphone. Source separation models have made real progress here and can sometimes pull voices apart usably. They are also generating audio, with the honesty consequences that carries — which matters in journalism and anywhere the recording is evidence.

Nothing recorded. The take with the microphone switched off. This is more common than anybody admits, and the fix is re-recording, not repair.

The resynthesis question

A newer class of tool runs a recording through a model and regenerates the voice rather than filtering it. The results can be remarkable — a phone recording from across a room returning close to studio quality.

Two cautions, and they are the same ones as elsewhere in this course. The tool is generating audio, so it can shift timbre and invent detail that was never present, which makes it inappropriate anywhere the recording is evidence, journalism or a record of what somebody said. And once a tool moves past cleanup into altering words or producing a voice, that is synthetic media, with labelling requirements that differ by country and are changing, and a consent requirement that is not optional.

The order of operations

Repair before mixing, and repair in this order: de-click → de-hum → de-noise → de-reverb → EQ → compression. Each earlier step removes something the later ones would otherwise amplify.

Today

Take your worst recording and apply exactly one thing: a high-pass at 100 Hz. Then apply six decibels of noise reduction, no more. Compare with the original. That is most of what is available, and it is usually enough.