Frequently asked questions

A peak reading reports the largest single sample or the largest reconstructed excursion in the piece. A loudness reading reports perceived level, averaged over time with a filter that resembles the ear and a gate that ignores silence. A master can peak twice and still be quiet, or peak once and sound loud, so the two figures answer different questions and are read together.

Choose the frame from the event, not from habit. A short event such as a click needs a short frame, in the 512 to 1024 range at a typical sample rate, so the event stays localised in time. A steady tone such as hum is best seen with a longer frame, 4096 or more, because the extra length resolves its frequency more precisely. See spectral analysis basics for the underlying trade.

Almost always because the subtraction stage removed high-frequency content along with the hiss, which lives in the same band. Raising the subtraction factor or placing the floor too high takes energy that was part of the wanted signal. Lower the factor, shorten the profile selection, or apply a gentle high-shelf lift after the repair and re-audition against the untouched reference.

Measure before limiting, limit for control, then measure again and trim to the target. If loudness is forced onto the target by limiting alone, the peaks are shaved to reach it and the piece loses transients that no later stage can return. A single gain trim after a light limiter reaches the same target with a far gentler result.

True peak measures the reconstructed waveform between samples rather than the samples themselves. A lossy encode can raise that waveform above the stored sample values, so a file that looks safe can arrive clipped. Reserving one decibel below full scale, written -1 dBTP, takes almost nothing in perceived loudness and removes the whole class of overs. Details are in LUFS mastering standards.

A working figure is three to six decibels of headroom on the loudest peak of the mix, with no true-peak overs anywhere in the chain. That margin lets the mastering stage make small corrective moves without running out of room, and it keeps any intermediate result well away from the ceiling until the final gain decision is made.

It matters whenever the delivery format is narrower than the working format. Truncating to a smaller step size without dither turns low-level detail into correlated distortion, heard as a gritty tail on fades. A small amount of shaped dither converts that distortion into a low, even noise and is added once, at the final step, never inside the working chain.

Use narrow notches, one per harmonic, and check the fundamental frequency exactly before filtering. A notch that is too wide removes neighbouring bass content and the result sounds hollow. Where the mains frequency drifts, a tracking filter follows it; where it is stable, a fixed comb of narrow notches is more transparent and easier to keep narrow.

Only where the flattened crest still carries a plausible shape, and only with a dedicated declip stage. A short run of clipped peaks in an otherwise healthy take can often be rebuilt from the surrounding slope. A long, sustained flat section carries no information about the lost peak, so any repair is an invention, and the honest course is to note the damage and work with what remains.

Long enough to average several hundred frames, which usually means two to five seconds of genuine room tone with no speech, no breath and no movement. A shorter selection stays noisy and the subtraction follows that noise into the piece. A longer selection is harmless provided nothing wanted appears inside it.

Usually because they disagree about what they measure. One may report sample peak while the other reports true peak, or one may use a sliding window while the other integrates over the whole piece, or one may apply a gate and the other may not. Read the label before the number, and compare two meters only when both state the same quantity and the same window.

Revisions become quick. Because the source region is read-only and every change is a parameter, a late decision is a curve edit rather than a re-record. The expense is a wider internal format, latency compensation for nodes that report it, and a graph that has to be evaluated in a sensible order. The article non-destructive effects sets out that trade in full.

Write to the editorial desk with the article name, the figure that appears wrong, and any source you can cite. Corrections are published with a dated change note so a reader can see what was revised and when.