Normalize audio means changing a whole recording’s level so it lands on a target, instead of wherever it happened to be recorded. One gain setting moves the entire file up or down. A whisper and a shout in the same file both move by the same amount, so the shape of the sound stays as it was and only the overall level changes.
People mean two different jobs by the word. They are worth keeping apart, because one is about the tallest peak in a file and the other is about how loud it sounds to a person.
Peak normalization: fit under a ceiling
Peak normalization finds the loudest point in the file and turns the whole file up or down until that point sits at a ceiling, usually between −1 and −3 dBFS. It is a single gain change, so it is instant and it adds no distortion.
The catch is that a peak is one sample, not an impression. If a chair scrapes once and the speech is quiet, the whole file gets turned down to fit the scrape and the speech stays quiet. Two files can both touch −1 dBFS and still sound nothing alike next to each other.
Loudness normalization: hit a level people hear
Loudness normalization measures the average loudness across the file, weighted the way hearing works, and turns the file until it reaches a target. The unit is LUFS. This is the version podcast apps, music services and broadcasters rely on, because it makes two files play at the same level one after the other.
The targets in common use are −23 LUFS for European broadcast (EBU R128), near −16 LUFS for podcasts, and about −14 LUFS for music streaming. A platform that normalizes will turn a loud file down to meet its own target, so a file far above it gains nothing.
| Peak normalization | Loudness normalization | |
|---|---|---|
| What it measures | the loudest sample | the average, as heard |
| Unit | dBFS | LUFS |
| What it guarantees | no clipping over the ceiling | files match in level |
| Where it is used | upload limits, quick fixes | podcasts, streaming, broadcast |
What normalizing does not fix
This is the part people miss. Normalizing is a level change and only that.
- Noise gets louder too. Hiss, hum, fan noise and traffic are in the file, so they rise with the voice. Remove the noise first, then set the level.
- Clipping stays clipped. Once the waveform is flattened at the top, the missing part is gone. Turning the file down does not restore it.
- It does not even out one speaker. If someone leans back from the mic, the whole file moves together and the dip stays where it is.
- It adds nothing. No detail, no warmth, no room correction.
When it is the right move
- A lecture or interview recorded on a phone, far quieter than anything else you publish.
- Two mics in one conversation that came in at different levels.
- A batch of episodes where some were recorded louder than others.
- A platform or a client that asks for a specific level.
Normalizing in Gleam Audio
Enhance is the tool for this. Since 2.2 it opens with a Kind of recording preset, Podcast, Lecture or class, Voice-over or narration and so on, which fills in the cleanup steps for that kind of sound. Target loudness is separate: pick where the audio is going, a podcast, an audiobook, music streaming or broadcast, and every file comes out at the level that place expects. It also clears the low rumble that mics pick up from desks and traffic.
Each file in the list shows how loud it is now, like “Now −18.4 LUFS, peaks at −1.1 dBFS”, so you can see how far it has to go before you run it. After 2.1 the same list has a Before and After switch, so you can hear the change at the same moment in the recording.
Nothing is uploaded. Enhance runs on your computer, your original file is left untouched, and it works on the whole list in one run. Every tool in the app takes video files too and works on the sound inside them.
Normalizing from the command line
ffmpeg has a filter called loudnorm that does EBU R128 loudness normalization. One pass measures and corrects as it goes:
ffmpeg -i talk.wav -af loudnorm=I=-16:TP=-1.5:LRA=11 talk-normalized.wavI is the integrated loudness target, TP the true peak ceiling, and LRA the loudness range it allows. ffmpeg also supports a double pass for files: the first pass reports the measurements, and the second applies them, which is more accurate on a long recording. For live audio, the single pass is the only option.
Measured here in a test on a quiet tone: the input sat at −43.8 LUFS with a true peak of −40.1 dBTP, and one pass at I=-16 returned an output of −16.0 LUFS.
The short answer
If the goal is for files to match each other, normalize loudness, and −16 LUFS is the usual target for speech. If the goal is to keep a file under a limit before upload, peak normalization is enough. Neither one replaces noise removal, and neither one rescues a recording that was captured badly.
Related term: LUFS. For a specific editor, see how to normalize audio in DaVinci Resolve.