LUFS Reference
platform targets and the ducking gap

Platforms turn loud uploads down rather than turning quiet ones up beyond a point. A master crushed to -9 LUFS is played back at -14 with all its dynamics squashed out, so it sounds flatter than a properly mastered file at the same perceived level.

Mix to the target, keep true peak at -1 dBTP to survive lossy encoding, and let the platform do nothing.

The single most useful number in this reference is the gap: music sits roughly 6 LU below dialogue while someone is speaking, and can come up to about 3 LU below in gaps. Absolute fader levels vary per track; the gap is what the ear judges.

Pair it with an 80 to 150 millisecond attack and 300 to 500 millisecond release and ducking becomes inaudible, which is the whole point.

Every major platform applies loudness normalisation on ingest. If your master is louder than the platform target, it is turned down; if it is quieter, most platforms leave it alone rather than boosting it. That asymmetry is the whole reason mastering hot is counter-productive: you do not get a louder video, you get the same perceived level with less dynamic range and a squashed, fatiguing mix.

The practical consequence for short-form is specific. A heavily limited edit that hits -9 LUFS gets pulled back to roughly -14 on playback, but the limiting damage stays. Dialogue that was riding against a brickwall now sounds flat next to a competitor's video that was mastered at target and left breathing room. On a muted-first platform this matters less for the first second and much more for the viewer who unmutes at second three.

The other side of the same problem is under-target masters. A -20 LUFS export is not normalised upward on most surfaces, so it plays noticeably quieter than everything around it in the feed. Viewers do not diagnose this — they scroll.

Integrated loudness is the average across the whole programme and is what platforms measure against. Short-term loudness is a rolling three-second window and is what tells you whether the hook is competitive in the first beat. True peak is the inter-sample maximum after conversion to lossy audio, which is why a -0.1 dBFS peak can still clip after encoding.

Measure integrated on the finished timeline, not on the dialogue stem. Measure short-term across the first five seconds separately — on short-form this is the number that decides whether your opener cuts through, and it is routinely 4-6 LU below the programme average on edits that feel weak despite measuring correctly overall.

Set the true peak ceiling at -1 dBTP for social delivery. Broadcast conventions of -2 dBTP are safe but leave audible headroom on the table; -0.1 dBTP invites codec-induced clipping on AAC playback.

Music beds mixed by eye rather than by ear against the voice. If the bed measures 8-10 LU below the dialogue it will still mask consonants during dense passages — ducking has to respond to the dialogue envelope, not sit at a fixed offset.

Multi-speaker interviews where each speaker was normalised independently to the same peak. Peak normalisation ignores perceived loudness, so a soft-spoken guest still sounds quiet next to a projecting host. Match on loudness units, then fine-tune by ear.

Voiceover recorded in a live room and then compressed hard to fix the level. Compression raises the room tone along with the voice; the fix is treatment or a closer mic, not more gain reduction.