Loudness normalization
Loudness normalization chooses gain from a programme measurement while accounting for the peak headroom and the intended playback target.
Loudness normalization reduces level jumps between programmes by measuring each programme and choosing a gain that brings it toward a common target. With constant gain, every part of the programme is scaled by the same amount. A quiet verse stays quieter than its chorus.
That is different from peak normalization, which aligns the highest peak, and from dynamic range compression, which changes gain as the programme develops.
Calculating the adjustment
For a programme measured in LUFS, the initial calculation is:
gain in dB = target loudness − measured loudness
amplitude multiplier = 10^(gain / 20)
A measurement of −18 LUFS and a target of −23 LUFS give −5 dB, or an amplitude multiplier of approximately 0.562. This reduces both the average level and the peaks by 5 dB.
The −23 LUFS target is the EBU R 128 broadcast reference used for this example. Playback systems can select another target. The target must be identified along with the measurement method; it is not a universal Emby setting.
Constant gain usually shifts the integrated value by the same number of decibels. A material change in which blocks cross the meter's absolute gate can alter that simple relationship, so processed output still merits measurement when an exact result matters.
A target and a peak ceiling can disagree
Suppose the output has a chosen true-peak ceiling of −1 dBTP, also matching the EBU production example for linear audio.
| Programme | Measured loudness | Measured true peak | Gain to −23 LUFS | Predicted peak |
|---|---|---|---|---|
| A | −18 LUFS | −1 dBTP | −5 dB | −6 dBTP |
| B | −26 LUFS | −2 dBTP | +3 dB | +1 dBTP |
A can simply be turned down. B would exceed the ceiling by 2 dB if the full +3 dB were applied.
With constant gain alone, B can receive only +1 dB before reaching −1 dBTP. Its resulting loudness is approximately −25 LUFS, short of the target. This is a valid outcome: a programme with prominent transients can remain quieter when the workflow preserves those transients.
A limiter could control the peaks and permit more gain. That would be a different processing choice, changing the waveform and potentially its dynamics. Some tools combine loudness targeting with dynamic processing; the label “normalization” by itself does not establish which approach was used.
For a playlist, lowering the target can make constant-gain matching possible for more tracks. For a movie, retaining more peak headroom may be preferable to repeatedly constraining effects. The practical choice follows the listening purpose.
The gain can live in metadata or in new samples
A scanner can store a measurement and recommended gain while leaving the media's samples intact. ReplayGain is one example. A supporting player reads that information and applies the gain during playback. Another player may ignore it or select a different track/album policy.
Alternatively, an offline process can multiply decoded samples, then store a new PCM or encoded file. That changes the media. Re-encoding to a lossy format also adds a new codec generation; merely writing a gain tag does not.
The distinction makes a useful Emby test concrete. Determine whether the selected app reads a tag, whether a server conversion applies gain, and whether the operating system or receiving device adds another level process. Support for a tag in a library scanner does not automatically imply that every playback route applies it.
The analyzed programme must match the delivered one
Channel conversion and track selection can change the loudness being reproduced. For example, adding a centre channel into a stereo downmix produces a new combination of signals; the original multichannel measurement does not directly describe that result.
A robust processing workflow therefore analyzes the intended programme, chooses the gain and peak policy, applies the processing, and checks the delivered output. In an offline workflow, analysis of the complete programme can inform one constant gain from its first sample. A live process has only the signal seen so far unless measurements arrive in advance.