📏 LUFS Loudness Standards: Getting AI Music Broadcast-Ready

Nobody's text-to-music model asks where you're publishing before it renders the file, which means nothing about the generation step targets a delivery loudness spec, because the model has no idea what the delivery spec is. That's not a knock on Suno or Udio or any AI music tool specifically. It's just a fact about the pipeline, and it means the loudness check that used to happen automatically in a mastering pass now has to happen deliberately, by you, after the fact, or it doesn't happen at all.

By the numbers#

• Streaming targets cluster at -14 LUFS integrated: Spotify, YouTube, Tidal, Amazon Music, SoundCloud, with Apple Music a notch quieter at -16 LUFS
• Broadcast and cinematic delivery under EBU R128 sits at -23 LUFS integrated, tolerance ±0.5 LU
• True peak ceiling on nearly every platform above: -1 dBTP
• The measurement itself comes from ITU-R BS.1770, a K-weighted algorithm that filters audio the way human hearing actually perceives frequency, then gates out silence before it reports one number

Two numbers, one confusing decade#

I cut audio for broadcast for twenty years before any of this AI-generation stuff existed, and the loudness spec is the one part of the job that got more standardized while everything else got less predictable. Back when everyone mastered to peak level instead of perceived loudness, stations and networks fought a genuine "loudness war": every advertiser wanted their spot to feel louder than the show around it, so everyone kept pushing peak levels up, and average perceived loudness crept up with it for years until regulators stepped in. EBU R128 and its American cousin ATSC A/85 exist specifically because peak-level mastering was a race nobody could win. LUFS measurement broke that race by measuring what a listener actually perceives, gated and K-weighted, instead of a raw peak number anyone could cheat by leaving more silence in a track. The practical point: if you're still mastering by eye on a peak meter instead of an integrated LUFS reading, you're solving a problem the industry already fixed thirty years ago, the hard way.

The gap AI tools don't close for you#

A voice model like the ones covered in our ElevenLabs guide renders audio optimized for clarity and naturalness at the waveform level, not for -14 LUFS, not for -23 LUFS, not for any delivery spec, because it has no concept of where the file is going next. Same story for AI-generated music: the model's job ends at "does this sound good," and loudness-to-spec was never part of that job description. If you paste a raw AI export straight into a video timeline or upload it straight to a platform, you're publishing whatever level the model happened to land on, and that level is not going to reliably hit -14 LUFS just because it sounds fine on your studio monitors at whatever volume you had them set to.

This matters more than it sounds like it should, because platforms don't all handle a mismatch the same way. YouTube's normalization only turns things down: it attenuates a hot master to -14 LUFS, but it will not boost a quiet one up to meet the target. Publish an AI-generated track that renders at -20 LUFS and it plays back quiet, permanently, relative to every properly-mastered video around it in someone's feed. That's a real, audible competitive disadvantage against channels that check their numbers, and it's invisible until someone tells you to turn their video up.

The actual fix, in one pass#

Run your export through a proper loudness analysis before you publish it, not after someone complains. ffmpeg's loudnorm filter does a full two-pass integrated-loudness and true-peak read for free, and it'll tell you exactly where your file sits against -14, -16, or -23 LUFS depending on where it's headed. Most current DAWs also ship an integrated LUFS meter standard now, so you don't need broadcast-suite gear to read the number. You need to actually look at it once before export instead of trusting your ears at an arbitrary monitor volume, which is the single most common way a mastering pass gets skipped without anyone noticing.

If you're assembling a podcast episode with AI-narrated segments and AI-generated music beds, the kind of mixed pipeline we walk through in our podcast production workflow, check both elements separately before you commit to a final mixdown. A voice track and a music bed generated by two different tools on two different days are not going to land at matching loudness by coincidence, and neither one is going to land at your platform's target by coincidence either. The same discipline applies to AI-generated sound effects layered into a mix; see our sound effects guide for the workflow side, since a hot one-shot SFX hit can trip a true-peak ceiling that the rest of your mix never comes close to.

Where I actually land on this#

I don't care how good a text-to-music model sounds in isolation until I've run the export through a meter and confirmed it survives contact with an actual delivery target. That's not skepticism for its own sake. It's the same test every broadcast mix had to pass before AI tools existed, and the fact that a model generates convincing audio says nothing about whether that audio is loudness-compliant for wherever you're about to put it. Meter it, then publish it. In that order, every time.