MLI – The Musical LUFS Initiative

Toward a More Musical Loudness Standard

Loudness normalization is one of the most important, and least emotionally neutral, parts of modern music playback. For listeners it is mostly invisible. A song starts, the platform adjusts its gain, and the next song follows at a comparable level. When it works, nobody notices. When it fails, it can quietly damage the musical relationship between songs in ways no artist, mix engineer, or mastering engineer ever intended.

The Musical Loudness Initiative — MLI — is an attempt to address that systematically, and to open a serious, practical discussion about whether a loudness model can be built that is more musically relevant than the one currently used across streaming platforms.

The problem with current normalization

Today’s loudness normalization is based on ITU-R BS.1770, a standard developed for broadcast television. The problem it was designed to solve was a specific one: jarring level jumps between TV programs and commercial breaks, where the primary material is speech. Music was part of that world, but it was not the main subject.

That standard now governs how music is normalized on every major streaming platform.

Integrated LUFS is far better than peak normalization and more meaningful than RMS. But it has real limitations as a universal rule for music. It measures the signal, not the intention. A sparse voice-and-guitar recording may be raised until it feels unnaturally large, while a dense, limited track is pulled down until it loses impact. A song with a quiet intro receives more playback gain than a track that stays intense throughout, meaning the song with more internal contrast can end up with its loud section pushed forward in ways the mastering never intended. LUFS also does not fully account for density: a wall-of-sound master and a whispered vocal track can be pushed toward the same target, regardless of how different their intended roles are.

Music has always used dynamics and density as expressive tools. A track meant to be subdued should feel subdued. A track meant to assault the listener should be allowed to do that. A loudness standard for music should not pretend that every track wants to occupy the same emotional space.

MLI

MLI is not a proposal to abandon LUFS or replace one rigid standard with another. It is an initiative to build on what exists — to find measurement approaches that incorporate more of the musical structure of a recording, rather than treating all program material as equivalent.

That discussion should not belong to one person. It should involve mastering engineers, mix engineers, artists, developers, researchers, and regular listeners. A loudness standard affects everyone who makes, releases, distributes, or listens to music. MLI is an open invitation to test, challenge, and improve the ideas it puts forward.

Emelie

Emelie is the practical test environment for MLI. It is not a mastering tool. It does not render files, write tags, or alter source material. It analyzes audio files, calculates a gain proposal for each loaded loudness method, applies that gain live during playback, and lets the listener decide which method produces the most musically coherent result — in blind or open listening mode.

In blind mode, methods are hidden behind anonymous columns and randomized between trials. The listener votes for the method that creates the most believable playback relationship, without knowing which is which. This matters because a familiar method name should not automatically receive trust, and a new method should not be rejected simply because it is unfamiliar.

Alongside its built-in methods, Emelie can load custom loudness methods from open .mlim files — JSON documents describing a measurement approach, without executable code. This makes method experimentation accessible to engineers, developers, and researchers who want to test their own ideas.

MLI Natural 1.5

MLI Natural 1.5 is the primary method developed within MLI and the one currently shipped with Emelie.

Rather than applying the same measurement rule to every track, it first analyzes the track’s musical structure and selects a profile. The profiles include:

• Ultra-dense / wall of sound

• Dense full-on / EDM

• Full-on / sustained intensity

• Intro then sustained loudness

• Short climax / crescendo

• Intentionally quiet

• Balanced fallback

Each profile represents a different type of musical behavior. A dense full-on track is not treated the same way as a song with one short climax. An intentionally quiet piece is not lifted to the same foreground position as a normal pop production that simply measures lower. A track with a quiet intro is not allowed to push its loud section forward just because the intro has pulled the integrated average down.

To select a profile, the method examines structural evidence in the signal: how loud the loudest sections are, how much of the track is active, how sustained the intensity is, how peaks relate to loudness, and how the loudest 10 or 30 seconds compare to the integrated value. It does not understand intention in a human sense. It uses measurable musical behavior as evidence.

MLI Natural 1.5 is not the final answer. It is the strongest method developed within MLI so far, included with Emelie so it can be tested, challenged, and used as a starting point for something better.

Open questions

Album Gain is one of the most important areas MLI has not yet fully addressed. For music released as a collection — an album, an EP, or a two-track single — Album Gain should be the default normalization principle, because it preserves the internal loudness relationships between tracks. Those relationships are part of the mastering.

Today, Album Gain is typically calculated by identifying the loudest measured track, normalizing it to the target, and applying the same offset to every other track. This keeps internal balance intact, but it also means one track controls the gain treatment for the entire release. An album containing one dense, full-on production alongside mostly voice and acoustic guitar may have the whole release pulled down because of that one track — including songs that were already sitting at a musically appropriate level.

A more coherent approach would be to measure the release as a whole and derive one shared gain value from the complete body of music, rather than letting a single outlier set the level for everything else. This is not part of Emelie’s current engine, but it is one of the most important next questions for MLI.

The target level is also open. If a musically aware method leaves sparse and intentionally quiet material lower than a strict integrated LUFS method would, it may become possible to use a slightly higher reference — perhaps around -12 or -13 LUFS — without making everything feel uniformly louder. The measurement method and the target interact, and that interaction needs listening tests and data. That is why Emelie exists.