Broadcast

Broadcasting solved the loudness problem with a measurement rather than a compressor: one number per programme, gated so that silence does not dilute it, and a range that says how much the programme moves.

The word loudness carries two meanings on this site, and they are not versions of each other. Here it is an energy measure: a K-weighted mean square over the whole programme, gated, reported in LUFS, designed so that two programmes normalised to the same number feel equally loud on the same playback chain. In Psychoacoustics it is a perceptual magnitude in sones, computed by an auditory model with masking and compression. A broadcast deliverable is specified in LUFS; a product-noise sensation is specified in sones. Reaching for the wrong one is the commonest mistake in this area.

The quantities are few. Loudness is reported in LUFS by the EBU and in LKFS by the ITU — identical units — and 1 LU is 1 dB, so a loudness difference and a level difference are the same size. EBU R 128 sets the delivery target at −23.0 LUFS with a true-peak ceiling of −1 dBTP. The loudness range, in LU, says how far the programme moves between its quiet and loud passages, which is what decides whether it needs dynamic treatment before normalisation.

Four documents own four different things, and the section is easier to read once that is clear. ITU-R BS.1770 defines the algorithm: the K-weighting pre-filter — a roughly +4 dB spherical-head shelf followed by the RLB high-pass — the mean square in 400 ms blocks at 75 % overlap, the channel-weighted sum, and the two-stage gate that makes the number usable on real programme (an absolute gate at −70 LKFS drops digital silence, then a relative gate 10 LU below the mean of the survivors drops the quiet passages that would otherwise dilute a dialogue level). EBU R 128 sets the target and the ceiling. EBU Tech 3341 defines the EBU Mode meter — the momentary, short-term and integrated time scales, and the compliance test set. EBU Tech 3342 defines the loudness range. True peak is measured on an oversampled signal because an inter-sample peak can exceed every sample value, so a file that reads −0.2 dBFS can still clip a converter.

Broadcasting standardised a second meter, forty years earlier and for the opposite question. ITU-R BS.468-4 measures the noise on a programme circuit rather than the programme: a weighting network peaking at +12.2 dB near 6.3 kHz, read by a quasi-peak detector, reported in dBqps. The two are worth reading side by side because they are built on opposite principles. BS.1770 is a gated mean square with its coefficients printed to fourteen digits; BS.468-4 is a peak follower whose dynamics are specified only by eleven tone-burst acceptance windows, with no time constant printed anywhere in it.

• Programme loudness (EBU R 128): the ITU-R BS.1770 K-weighting, gated 400 ms blocks and channel-weighted sum, the EBU R 128 target and ceiling, the Tech 3341 momentary, short-term and integrated meters, the Tech 3342 loudness range, the Annex 2 oversampled true peak and the Annex 3 channel weights for advanced sound systems — validated against the EBU test signals and ending in an EBU R 128 report fiche.
• Quasi-peak programme meter (ITU-R BS.468-4): the psophometric noise meter — the clause 2 quasi-peak detector, its eleven Table 2 and Table 3 acceptance windows, the clause 2.6 calibration that makes a reading dBqps, and the three fitted time scales those windows identify only to within a factor of 1.61 to 2.09.

• BS.1770-5 Annex 4, object-based audio, is out of scope, and the library implements no spatial renderer, so an object-based programme has to be rendered to a loudspeaker layout before any of this applies. EBU Tech 3343 is cited as production practice around these numbers, not as an algorithm: nothing here runs it. And loudness normalisation itself — the gain change, and any limiting that follows it — is a production step this library does not perform: it measures the programme and tells you the offset, and applying it is your encoder’s job.