Loudness Normalization (LUFS and True Peak)

Loudness normalization is the practice of measuring how loud a piece of audio sounds to a listener, as a single number, and adjusting it to a target so that programs play back at a consistent level. The measurement unit is LUFS (loudness units relative to full scale; LKFS, the term used by the ITU and US broadcast standards, is the same measure), defined by the ITU's BS.1770 standard. Broadcasters adopted loudness targets to end the era of commercials that blasted louder than the shows around them, and streaming services adopted normalization so listeners don't have to reach for the volume control between tracks. For anyone delivering video or podcasts to YouTube, streaming services or broadcasters, the practical questions are what target to hit, and which published targets are real.

How Loudness Is Measured

A level meter shows how strong a signal is; it doesn't show how loud it sounds. Human hearing is less sensitive to low bass and more sensitive to the upper midrange, and a short burst sounds less loud than a sustained tone at the same level. ITU-R BS.1770, now in revision 5 (November 2023), defines a measure that accounts for this. It applies a frequency weighting called K-weighting, averages the energy across all channels, and uses gating to ignore silence and very quiet passages, so that a program with long pauses isn't scored as quiet. The same standard defines how to measure true peak.

Three numbers matter in practice:

• Integrated loudness (LUFS): the gated average over the whole program. This is the number targets refer to.
• True peak (dBTP): the highest level the waveform reaches, including peaks between samples that appear when audio is converted back to analog or re-encoded to a lossy format. Sample-peak meters miss these.
• Loudness range (LU): how much the loudness varies across the program, a measure of dynamics.

Broadcast Standards

Broadcast regulators turned the measurement into rules. In Europe, EBU Recommendation R 128 (version 5.0, November 2023) sets a target of −23 LUFS, with a tolerance of ±0.5 LU for most programs (±1 LU for live), and a maximum true peak of −1 dBTP; companion documents define the meter (EBU Tech 3341), loudness range (Tech 3342) and production and distribution practice (Tech 3343 and 3344). In the United States, the CALM Act of 2010 led the FCC to require broadcasters to follow ATSC A/85, whose target is −24 LKFS, so that commercials play no louder than the programs around them. Streaming services that commission content set their own specifications; Netflix's delivery specification for branded content, for example, measures dialogue-gated loudness at −27 LKFS.

Platform Targets: Official Versus Folklore

Online, the picture is messier, and it is worth separating what platforms publish from what the community has measured.

• Spotify (official). Spotify publishes its normalization: −14 LUFS by default, with listener-selectable "Loud" (−11 LUFS) and "Quiet" (−19 LUFS) settings, measured with BS.1770. It advises keeping true peaks below −1 dBTP, or below −2 dBTP for masters louder than −14 LUFS.
• Apple Podcasts (official). Apple's audio requirements specify −16 LKFS, ±1 dB, with true peak no higher than −1 dB, measured per BS.1770-5.
• AES (recommendation). The Audio Engineering Society's TD1008, later standardized as AES77-2023, recommends about −18 LUFS for speech and mixed content and −16 LUFS for music in streaming and network distribution.
• YouTube (community-measured). YouTube doesn't publish a loudness target. Measurements by engineers suggest it turns down content louder than roughly −14 LUFS, and its "Stats for nerds" panel shows "content loudness" as the difference from its reference, but it doesn't turn quiet videos up. The commonly quoted −14 LUFS figure is folklore backed by measurement, not a specification.
• Apple Music (community-measured). Apple documents Sound Check but not a numeric target; the commonly quoted figure of about −16 LUFS comes from measurement.

Two points follow. First, when a platform normalizes, mastering louder than its reference gains nothing: the platform turns the file down, and the only lasting effect is lost dynamic range and potential distortion. Second, when a platform doesn't publish a target, any number quoted for it is a measurement of current behavior, which can change without notice. A widely used compromise for online and short-form video is to deliver around −14 LUFS integrated with true peak at or below −1 dBTP, while understanding that this is a convention, not a specification.

Loudness in Automated Pipelines

Loudness is one of the easiest things to get wrong in automated video production, from AI video editing tools to fully agentic pipelines, and one of the easiest to check. Generated voiceover from text-to-speech, interview audio cleaned up with speech enhancement and generative music each arrive at different levels; a pipeline that concatenates them without measurement produces jumps a listener notices at once. Two practices help. Level speech from measured loudness rather than a fixed gain, so each speaker and each take sits at a consistent level, and lower music under dialogue with ducking. Then measure the finished render, not the plan, as part of automated video QC: integrated loudness against the platform target with a tolerance of one or two LU, and true peak against −1 dBTP. A program that is on target but peaks above 0 dBTP will clip on many playback paths.

FFmpeg's ebur128 and loudnorm filters implement the EBU measurement and a normalization pass, which makes this check available to any pipeline without commercial tools.