Peak vs loudness normalization: making clips sound equally loud

Digital audio samples are numbers between −1 and +1. Zero dBFS — decibels relative to full scale — is the largest value the format can hold, so every real level is a negative number: −6 dBFS is half the amplitude, −20 dBFS is a tenth, −60 dBFS is a thousandth and close to inaudible on a normal listening level.

Peak level is simply the loudest single sample in the file. It tells you how close you are to clipping and nothing else. RMS — root mean square — is the energy average over a window of time, and it tracks perceived loudness far better, because loudness is about sustained energy rather than one transient. A recording with long silences between sentences has a low RMS and can still have a peak at −1 dBFS.

LUFS is the standardised version of that idea: ITU-R BS.1770, adopted in Europe as EBU R 128. It is an RMS measurement with K-weighting — a filter that approximates the ear’s frequency response and de-emphasises deep bass — plus gating that ignores the quiet passages so a long pause cannot drag the number down. One loudness unit equals one decibel. AudioCutlet measures plain RMS without weighting or gating, which lands within roughly a decibel of a LUFS reading on speech and typical music, and drifts further on very bass-heavy material.

Peak normalization finds the loudest sample and multiplies every sample by whatever factor puts that one on the target — −1 dBFS in this editor. It is a single multiplication, so it never distorts and it is perfectly reversible: the relationship between the quiet and loud parts is untouched.

Which is exactly why it cannot make two clips sound equally loud. Peak-normalize a sparse acoustic guitar recording and a heavily compressed pop master to the same −1 dBFS and the pop track will sound roughly ten decibels louder, because its average level sits much closer to its peak. The meters agree; your ears do not.

Use it when you want guaranteed headroom before export, when a recording came out too quiet and you want the full range back without changing its character, or when the file is going somewhere that will run its own loudness pass. It is the safe default, and it is what the Normalize button does unless you switch it to RMS.

Loudness normalization aims at an average instead of a peak, which is why it can actually match clips. For spoken word, −16 LUFS is the widely used figure for stereo podcasts (Apple’s guidance is −16 LUFS with true peaks no higher than −1 dBTP), and mono deliverables are often specified a few units lower. European broadcast is stricter and quieter: EBU R 128 sets −23 LUFS.

Streaming services normalize on playback rather than asking you to. Spotify plays back at around −14 LUFS, YouTube lands close to the same figure, and Apple Music’s Sound Check sits nearer −16. The practical consequence is that mastering louder than the target buys nothing: the service simply turns the track down, and you are left with the reduced dynamic range you traded for a loudness that no longer exists.

For the everyday job of making three clips sit together in one episode, the absolute number matters much less than using the same one for all three. −16 dB RMS is a good default for voice, −14 dB RMS for music; pick one, apply it to every clip, and they will match.

Suppose a clip measures −26 dB RMS and you ask for −16. That needs +10 dB of gain. If the loudest moment in the clip already sits at −6 dBFS, adding 10 dB would put it 4 dB past full scale — and there is nowhere past full scale to go. The samples get clamped flat, which is clipping: flat-topped waveforms, harmonic distortion, and an audible crunch on exactly the loudest hits.

AudioCutlet refuses to do that. It raises the gain only as far as keeping every peak under −0.1 dBFS allows, applies that amount, and tells you in plain words how much it applied and that it stopped short. The clip is then as loud as it can be without damage, which is usually the honest answer to the question you were really asking.

Closing the remaining gap is a job for compression or limiting: reduce the peaks first, then lift the whole signal. This editor deliberately has neither, because a carelessly set limiter does more damage to a recording than being two decibels quieter than the target ever will. If you truly need it, a compressor in a DAW is the right tool; otherwise peak normalization to −1 dBFS is the safe alternative.

Samples are points, and the continuous waveform reconstructed between them can rise higher than any of them. A file whose highest sample is exactly 0 dBFS can therefore produce inter-sample peaks above 0 dBFS in a digital-to-analogue converter, or after a lossy encode, because the decoder’s output is never bit-identical to its input.

That is what “true peak” measures: the level of the reconstructed waveform, computed on an upsampled copy and quoted in dBTP. It is why almost every delivery specification asks for −1 dBTP rather than 0, and why leaving a decibel of headroom is free — nobody has ever noticed a master being one decibel quieter.

AudioCutlet’s peak normalization targets −1 dBFS for exactly this reason. It measures sample peaks rather than true peaks, so with heavily limited material — modern pop masters especially — it is worth setting a slightly lower target or exporting WAV and checking in a metering tool if the file has to pass a broadcast spec.

Normalize last. Every earlier edit changes the measurement: removing a loud cough changes the peak, removing the silence between sentences raises the RMS, and a fade-out lowers it again. Trim first, then remove silence, then apply fades, then level — that order means the number you measure is the number the listener gets.

Do not skip the fades. A cut in the middle of a waveform is a step change in the signal, and a step is a click. Ten to thirty milliseconds at each end of a clip removes it inaudibly; a musical fade-out is usually half a second to two seconds. In this editor [ and ] apply the fade in and fade out with the length set in the operations panel.

Then listen, once, at a moderate level, on something resembling the device your audience will use. Meters catch clipping and mismatched averages; they do not catch a clip that is technically correct and still sounds wrong next to the one before it.