Two people on one podcast: one mic runs hot, the other is timid, and the mix is a roller coaster. Someone suggests “just normalize it.” That advice is half right, and the half that is wrong is the expensive half.
Two jobs, one name
“Normalize” describes two operations that measure different things:
| Peak normalization | Loudness normalization | |
|---|---|---|
| Measures | highest sample (dBFS) | perceived loudness (LUFS) |
| Does | applies one fixed gain | adjusts to a loudness target |
| Keeps dynamics | yes, exactly | yes, but the target is perceptual |
| Typical target | -1 dBFS peak | about -14 LUFS (streaming) |
Peak normalization asks one question: how much room is left before the signal clips? Then it moves the whole clip up or down until the answer is “none.” Loudness normalization asks how loud this feels and uses K-weighted measurement to match how ears hear mid frequencies. Streaming platforms use the second kind. They measure your track and turn it down to roughly -14 LUFS, so a louder master does not play louder; it just gets squashed harder.
What normalization cannot fix
- Uneven levels inside one clip. A single gain value moves everything. If speaker A is 8 dB hotter than speaker B, normalizing the file keeps that gap.
- A noisy floor. Amplifying the voice amplifies the hiss by the same amount. Normalize a noisy take and you get louder noise.
- Dynamics worth keeping. A whisper stays a whisper, which is good, but a shout without a compressor stays a shout.
When to do it, and in what order
- Split per speaker or per segment, and match clips with clip gain before any processing.
- Clean up first: high-pass, de-hum, gentle denoise.
- Normalize or set loudness last, after compression and limiting have done their work.
- Leave headroom: -1 dBFS peak for anything headed to MP3 or AAC, because lossy encoding can push intersample peaks past 0 dBFS and clip.
Do it when you are assembling multiple sources, publishing to a platform that normalizes anyway, or handing files to someone whose playback chain you do not control. Skip it mid-edit in a 32-bit float session — there it buys nothing.
Set the target by destination, not by habit: -1 dBFS peak for files you will encode, about -14 LUFS for streaming, and a re-record when the gap is between two speakers rather than two files.