Three measurements, three questions
Digital audio is stored as a sequence of sample values. It is tempting to find the largest one, call that “the level,” and treat every volume decision as a distance from zero. That number is useful, but it answers only one question: how close is the most extreme stored sample to the format’s numeric limit?
Audio delivery asks two other questions often enough that the names must stay separate:
| Measurement | What it examines | The question it answers |
|---|---|---|
| Sample peak | Discrete values stored in the file | Will multiplying these stored samples exceed full scale? |
| True peak | An oversampled estimate of the reconstructed waveform | Can playback or later processing exceed the peak ceiling between samples? |
| Programme loudness | Frequency-weighted energy integrated across time | How loud is the programme likely to be perceived relative to another? |
None is a more sophisticated synonym for another. A tool that calculates one should not label its result with the unit of another.
Sample peak: the boundary visible in the file
Integer PCM has a finite code range. A 16-bit sample runs from -32,768 to 32,767; floating-point WAV normally treats -1 to +1 as full scale even though the representation can store values outside it. A sample-peak scan converts those codes to a fraction of full scale and keeps the largest magnitude.
For a constant gain, the arithmetic is direct:
multiplier = 10^(gain dB / 20)
new sample = old sample × multiplier
If a largest positive sample is 0.5 and the lowest negative sample is -0.4,
the positive side allows a multiplier of 1 / 0.5 = 2, while the negative side
allows 1 / 0.4 = 2.5. The smaller factor is safe for both, so the largest
sample-safe boost is 20 × log10(2) = 6.0206 dB.
That is exactly the guarantee a local WAV volume adjustment can make after scanning the complete data chunk. It can also account for integer PCM’s slight asymmetry: the negative range contains one code whose positive counterpart does not exist.
Before you read on
Two recordings both peak at -1 dBFS. Must they have the same programme loudness?
No. One recording can contain a single short transient at -1 dBFS and spend the rest of its duration quietly; another can remain close to that ceiling for minutes. Their maximum stored value is equal, but their time-distributed, frequency-weighted energy is not. Peak is a boundary check, not a loudness model.
True peak: what happens between stored values
A digital-to-analogue converter does not play each sample as an isolated stair step. It reconstructs a continuous, band-limited waveform. That curve can rise above both neighbouring sample points, producing an inter-sample peak. A file whose stored samples remain below full scale can therefore overload a later converter, sample-rate conversion or codec stage.
ITU-R BS.1770 treats true-peak measurement as a distinct algorithm and describes oversampling used to estimate that reconstructed maximum. Its unit is commonly written dBTP. A sample-peak tool does not oversample, so it cannot honestly certify a dBTP ceiling. Leaving extra sample headroom may reduce risk, but only the measurement named by the delivery specification proves compliance.
This also explains why a codec pass can reveal a different peak. Filtering and reconstruction change the continuous waveform even when the decoded result sounds transparent. Peak checking belongs after the last relevant processing stage, not merely on the original WAV.
LUFS: energy, frequency and time
Two programmes can share an identical sample peak and sound dramatically different. A sparse interview may touch full scale for one consonant; a dense music master may remain near it continuously. Peak sees one extreme value and cannot distinguish those cases.
Programme loudness algorithms weight frequency, combine channels and integrate energy over a defined interval. ITU-R BS.1770 specifies the measurement basis; the EBU’s loudness work applies it to operational normalisation and distinguishes momentary, short-term and integrated views. LUFS is therefore a measurement of a programme over time, not another way to print a sample code.
A constant gain can move an already measured programme toward a LUFS target, but the loudness must be measured with the specified algorithm and measured again afterward. A sample scan alone cannot derive it. Gating can also change which quiet intervals participate, so a naive “add the LUFS difference” rule is not a universal compliance proof.
Why one control cannot promise all three
Consider a dialogue file with a single door slam. Peak normalisation may let the slam determine the gain and leave the dialogue quiet. Loudness normalisation may raise the dialogue substantially, then require separate true-peak management so the slam stays inside the delivery ceiling. A time-varying limiter could control the transient, but it changes dynamics rather than applying one constant factor.
Those are three different editorial choices:
- Constant gain preserves relative dynamics and channel balance.
- Loudness normalisation targets a measured programme level.
- Limiting or compression changes gain over time to control peaks or dynamic range.
The smallest honest tool exposes one of them and names the boundary. A WAV gain adjuster can apply constant gain and prevent new stored-sample clipping. It should say explicitly that it is not a true-peak limiter and not a loudness normaliser.
Trimming changes the measurement window
Cutting a PCM or float WAV at complete sample frames does not alter any retained sample. Even so, measurements made over the file can change. Remove the only transient and the new file’s peak falls. Remove a long quiet introduction and its integrated loudness may rise because the measured interval is different.
That is why container inspection, frame trimming and gain adjustment belong as separate operations. Inspect first to learn the actual format and duration; trim when the timeline should change; adjust gain when sample amplitude should change. After the edit chain is complete, run the true-peak and loudness meters required by the destination rather than carrying old analysis metadata forward.
A delivery checklist
Before changing a production file, ask for units and limits instead of accepting the word “normalise” by itself:
- Is the target a stored sample peak in dBFS, a true peak in dBTP, an integrated loudness in LUFS, or a combination?
- Which standard and version defines the measurement?
- Is one constant gain required, or may dynamics processing change gain over time?
- Must embedded peak, loudness or checksum metadata be recalculated?
- Does measurement happen before or after sample-rate conversion and lossy encoding?
For an ordinary local edit, a sample-peak scan is a transparent and useful safety boundary. For broadcast, cinema or a platform delivery, it is one input to the decision—not a substitute for the true-peak and programme-loudness measurements the specification names.
Common questions
Can audio clip even when every sample is below 0 dBFS?
Yes. Playback reconstructs a continuous waveform from discrete samples, and that curve can peak between the stored values. A true-peak meter estimates those inter-sample peaks; a maximum-sample scan does not.
Is peak normalisation the same as loudness normalisation?
No. Peak normalisation gives one extreme sample or estimated waveform point a chosen ceiling. Loudness normalisation measures weighted energy across time, so two programmes with the same peak can require very different gain to reach the same LUFS target.
Does a 6 dB gain make audio sound twice as loud?
A gain of about 6.0206 dB doubles waveform amplitude. Perceived loudness depends on spectrum, duration, listening level and the programme itself, so doubling amplitude is not a universal promise of sounding twice as loud.
Does trimming a WAV change its peak or loudness?
The retained samples stay unchanged, so their individual amplitudes do not move. But removing a section can remove the original highest sample and changes the time interval over which integrated loudness is measured, so file-level peak and LUFS results can change.
Which measurement should a broadcast or streaming delivery use?
Use the exact specification supplied by the recipient. Professional delivery commonly names an integrated loudness target and a maximum true peak together; a simple sample-peak adjustment cannot certify either requirement unless the specification explicitly asks for sample peak.