Why recordings differ in loudness and how to measure it properly
What LUFS is, which loudness levels apply in television, radio, streaming and phone announcements, and how to measure your own files.
Anyone who works with audio knows the moment. The ad break comes in louder than the programme it interrupted. A phone announcement is barely audible, and the on-hold music right after it makes the caller pull the phone away from their ear. Material from three different suppliers, each done properly, yet put together it sounds like a random collection.
This is not carelessness. It is the result of decades of measuring loudness in a way that did not match how people actually hear.
Why a peak meter misleads you
A classic meter shows the peak value, the single highest excursion of the signal. That tells you whether the file will clip. It does not tell you how loud it will sound.
Take two files normalised to the same peak value, say -1 dB. The first is a calm narration with natural dynamics: individual consonants hit the peak, but the average energy of the whole is low. The second is a dense, heavily compressed advertising spot in which almost all the material sits just below the peak. The meter will show the same number for both. The ear will hear a difference of well over ten decibels.
There is a second problem: the ear is not equally sensitive to all frequencies. The same energy level at 3 kHz is perceived as clearly louder than at 100 Hz. A peak meter knows nothing about this.
Setting levels "by ear" fails too, for a reason that is hard to work around. Hearing adapts to the current level within tens of seconds. After an hour of working on material, "even" means something different than it did at the start of the session.
What LUFS is
The solution turned out to be a measurement that models the way we hear. It is described by the ITU-R BS.1770 recommendation, and the unit is LUFS, or Loudness Units relative to Full Scale.
The measurement works like this: the signal passes through a filter that mirrors the sensitivity of human hearing across frequencies, then the average energy is calculated, and silent passages are excluded from that average so that pauses do not drag the result down. The outcome is a single number describing the loudness of the entire piece.
And that number has the property that was missing before: two files with the same LUFS value sound equally loud. Regardless of whether it is narration, an advertising spot or a piece of music.
Alongside the averaged value, two more things are measured. True Peak (dBTP) is the actual peak of the signal after conversion to analogue audio, usually slightly higher than the peak visible in the digital file. LRA (Loudness Range) describes the spread between the quietest and loudest passages, which tells you about the dynamics of the material.
Which levels apply where
This is where the most common misunderstandings with clients begin. There is no single universal level. There are standards tied to distribution channels, and they differ significantly.
The difference between -23 and -14 LUFS is nine decibels, which is a lot. It does not stem from a dispute over methodology but from listening conditions. Television is watched in a quiet room, where wide dynamics are an asset and the viewer sets the volume once for the evening. Streaming is listened to on headphones, on a tram, on the move, where material with wide dynamics simply disappears into ambient noise.
Phone announcements are a separate category. None of the standards above applies here, because the telephone path has a narrow bandwidth and its own level limits on the PBX side. The practical rule is the same, though: all messages in one system, together with the on-hold music, should be matched in loudness. The caller should not have to adjust the volume mid-call.
How to check your own files
Measurement is widely available today and does not require expensive equipment.
Every serious audio editing package has a built-in meter compliant with BS.1770. In Pro Tools, Reaper, Adobe Audition or DaVinci Resolve you simply switch it on and play the material in full, because the averaged value is calculated for the whole file, not for a fragment.
There are also free plugins that measure all three parameters at once, as well as command-line tools. In ffmpeg, the loudnorm filter can both measure a file and normalise it to a target value, which is useful for batch processing.
One thing is worth remembering: loudness normalisation is a change of gain across the whole file by a fixed amount, not compression. The dynamics of the material stay untouched. If a recording is meant to breathe, it will still breathe, just at a different level.
When there are more than a few files
For a single recording, measuring and correcting takes a minute. The problem appears with sets: dozens of messages for a phone system, a series of spots, an audio guide library, a set of backing tracks from various sources.
The only sensible approach then is batch processing against one target value. You pick the level appropriate for the distribution channel, run all the files through it, and you can be sure none of them will stand out from the set. This applies equally to voice-over material and to music. In the latter case, the problem is familiar to every musician playing to backing tracks, where files come from different producers and different years: we covered it in more detail in the context of normalising a whole set of files to one level in our CountIn tool.
The principle is shared: loudness is set once, for the whole set, according to the standard that applies wherever the material will end up.
What this means if you are ordering a recording
Three practical conclusions.
First, it is worth specifying the distribution channel when placing the order. Material prepared for television broadcast and the same material for a streaming channel are two different files, not one file in two copies.
Second, if the recording is to be edited together with other elements, for example voice-over plus background music plus a jingle, the level of the whole is set after mixing, not on individual tracks. The sum of three correctly set elements does not give a correctly set whole.
Third, with larger sets it is worth asking what value the material was normalised to. It is a single number that can be checked in a minute, and it saves a week of correspondence about something being "too quiet".
At Powitania.pl every recording goes through this stage as standard, and we match the level to the intended use of the material. If you are not sure which standard applies to your project, write to us before the recording, not after you receive the files. It is usually a matter of one sentence in the order.