Audio normalization explained: peak, RMS and loudness, and why streaming turns you down

"Normalize" is one button with two very different meanings. The old one lines up the loudest peak; the new one lines up how loud a track sounds, which is what streaming services and broadcasters use to make everything play at the same level. Knowing which is which explains why your carefully maximised track gets turned down on Spotify, and the audio normalizer offers both.

Peak normalization: the simple kind

Find the single loudest sample in the file, work out how far below full scale (0 dBFS) it is, and raise the whole track by that amount so the peak just touches the target (usually −1 or 0 dBFS). It never clips, never changes the dynamics, and it is what "maximise" and older normalize buttons do. Its weakness: a track with one loud snare hit and a quiet verse gets no louder than the snare allows, while a compressed track with no peaks gets much louder. Two peak-normalized files can sound wildly different in level.

Loudness: what ears actually hear

Perceived loudness depends on average energy over time, weighted for the frequencies ears are sensitive to, not on the highest peak. RMS (root mean square) was the first approximation; the modern measurement is defined in the ITU BS.1770 standard and reported in LUFS (loudness units relative to full scale). It applies a filter that mimics hearing, averages over the programme, and gates out silence so pauses don't drag the number down. Integrated loudness is the whole track's figure; short-term and momentary values track it over seconds.

LUFS and the streaming standards

Broadcast standards (EBU R 128 in Europe, ATSC A/85 in the US) fixed programme loudness at −23 or −24 LUFS. Streaming services normalise on playback to their own targets: Spotify around −14 LUFS, YouTube about −14, Apple Music −16, and podcast platforms −16 to −19 for spoken word. If your track is louder than the target it is turned down; if quieter, some services turn it up (with a peak limit) and some don't. The practical consequence is the important one: mastering hotter than the target buys nothing, and mastering with crushed dynamics to hit a big number just produces a flat track that is then turned down to the same level as everyone else's.

The loudness war, and how it ended

From the 1990s, releases competed to be louder on the radio and in playlists by limiting harder, until masters had almost no dynamic range and audibly distorted — the loudness war. Loudness normalization ended it: once every track plays at −14 LUFS regardless, the dynamic master sounds better than the crushed one at the same volume. Peak-to-loudness ratio (how much the peaks rise above the average) is now something to preserve, not eliminate.

Setting levels in practice

• Podcasts and voice: −16 LUFS integrated (−19 for mono), true peak below −1 dBTP. Consistent between episodes matters more than the exact number.
• Music for streaming: master to taste around −14 to −10 LUFS with true peaks under −1 dBTP; anything hotter is turned down anyway.
• Multiple files that should match (a course, an album): loudness-normalize them to one target, not peak-normalize.
• A single quiet recording: peak normalization to −1 dBFS is the safe, quick fix; loudness normalization if it must sit beside others.

Clean the recording first with the noise reducer (How audio noise reduction works, and why voices go watery), then normalize, then encode; the bitrate needed is covered in How MP3 compression works, and which bitrate you can actually hear.