A TEST METHOD FOR
LIVE LOUDNESS ADJUSTERS

CONTENTS

• 1. What is tested
• 2. The synthetic signals
• 3. The synthetic cases
• 4. Programme material
• 5. Running a case
• 6. What is measured
• 7. The reference processings
• 8. The criteria
• 9. What a report states
• 10. What the method does not cover

1WHAT IS TESTED

The device under test takes audio in blocks, adjusts its level toward a declared loudness target, and hands it back after a fixed delay. Everything it decides comes from audio it has already received, plus whatever look-ahead that delay buys it. That is what separates it from a two-pass normaliser, which reads the whole programme before it writes any of it.

Before a run, the device's declarations are fixed and stated in the report:

• Target T, in LUFS.
• True-peak ceiling, in dBTP, if the device has one.
• Latency L, in samples, at each sample rate used.
• Maximum boost, in dB, if the device limits how far it raises a quiet signal.
• Block sizes the host feeds, any block size the device declares to give output identical to the reference block's, and every setting that differs from the device's default.
• The software version, so a result can be tied to what was measured.

2THE SYNTHETIC SIGNALS

Every synthetic case is 48 kHz, two channels, generated in double precision. A case is a sequence of parts, each one signal at one level for a stated time. Four signals are used.

The random number generator

xorshift64*, with its state initialised from the seed s through splitmix64, because xorshift started directly on small consecutive seeds gives correlated first outputs. All arithmetic is unsigned 64-bit, wrapping.

init:   z = s + 0x9E3779B97F4A7C15
        z = (z ^ (z >> 30)) * 0xBF58476D1CE4E5B9
        z = (z ^ (z >> 27)) * 0x94D049BB133111EB
        x = z ^ (z >> 31)

next:   x ^= x >> 12;  x ^= x << 25;  x ^= x >> 27
        r  = x * 0x2545F4914F6CDD1D
        u  = (r >> 11) * 2^-53          in [0, 1)
        w  = 2u - 1                     in [-1, 1)

Seeds. Each family of cases has one seed (section 3). Part k of a case, counted from 0, uses seed + 16k, and channel c adds c. So every case in a family carries the same audio in the same place and differs only in what its name says.

Pink noise

White samples w from the generator, filtered by Paul Kellet's refined pink filter, in double precision, with all seven states at zero at the start of each part. The two channels are independent (each has its own seed).

b0 = 0.99886 b0 + w * 0.0555179      b4 = 0.55000 b4 + w * 0.5329522
b1 = 0.99332 b1 + w * 0.0750759      b5 = -0.7616 b5 - w * 0.0168980
b2 = 0.96900 b2 + w * 0.1538520      pink = b0 + b1 + b2 + b3 + b4 + b5 + b6 + w * 0.5362
b3 = 0.86650 b3 + w * 0.3104856      b6 = w * 0.115926

The filter was designed at 44.1 kHz, so its slope shifts slightly at 48 kHz. The signal is defined by the recurrence above, not by a claim about its spectrum.

Room tone, bursts and silence

• Room tone is pink noise as above. Its level is absolute, in LUFS, and does not move with the target.
• Bursts are white noise w multiplied by exp(-t / 0.060) for 200 ms, then silence, repeating every 500 ms. They are drawn from channel 0's generator and copied to both channels. One value is drawn per sounding sample; the silence between bursts draws nothing. A drum-like onset with a high crest factor.
• Silence is digital zero.

Levels

Each part is scaled so that its own integrated loudness, measured alone to ITU-R BS.1770-4, equals its stated level. Levels are written relative to the target, as T or T - 10, except room tone. Measuring once at a fixed scale and then scaling linearly is exact, because both gates are relative to the signal once it clears the -70 LUFS absolute gate. For a burst part, the measurement covers the whole part, silences included, as the gates treat them.

The set is defined for T = -23 LUFS. No part is louder than T + 10, so no signal needs headroom above digital full scale there. At a higher target the loudest parts would clip, so a report at another target states which synthetic cases it re-levelled or left out.

3THE SYNTHETIC CASES

Forty-five cases in eight families. Parts are listed in order, as signal, level, length. The need of a part is the gain that would put it exactly on target: T minus its level.

4PROGRAMME MATERIAL

Published reference files

Three files that the EBU and the ITU publish for testing loudness meters serve as programme anyone can obtain. The checksums identify the exact files.

13cb1c1ca8496c868a768bd1afc5e6a75841235000b68285e895bce62c2bdc6a  seq-3341-2011-8_seq-3342-6-24bit-v02.wav
3ff26c997d838aff36b4319f9e0a673c0b51fe26722ecb0153887b66f38c296f  1770-2 Conf Stereo VinL+R-23LKFS.wav
f8b318474b158c9f2ee842f0cfadf58ffcf504ffe8daa172220190d12dd0785b  1770-2 Conf Mono Voice+Music-23LKFS.wav

P5, the reference medley, is the three joined end to end in the order above, with no gap. The mono file is read as two channels, each carrying the mono signal at 1/√2, which keeps its BS.1770 loudness. The medley also carries the S3 cases. P4.2 is the same mono file as one channel.

Material of your own

A report may add programme of its own choosing, and should: speech, music, film, live streams, and joins between them. For each item it states the source, the licence, the length, the source's sample peak, and whether it has been through a lossy codec. Two limits follow from the criteria. An item shorter than 16 minutes cannot carry criterion G2, and a lossy item is left out of G3, because a codec's own inter-sample overshoot cannot be told apart from the adjuster's.

5RUNNING A CASE

• Start each case on a fresh instance, or immediately after a reset, with every declared setting applied before the first sample.
• Feed the declared block size. The reference block is 10 ms: 480 frames at 48 kHz, 441 at 44.1 kHz. The block-size check (G5) repeats every case at 20 ms and at 256, 512, 1024 and 2048 frames.
• End by feeding digital silence after the last sample until the delayed tail has come out.
• Align by dropping the first L output frames, so that output frame i is the same instant as input frame i. Every measurement below is taken on the aligned output.
• Check the latency rather than trust it: in S6.1 the first burst starts at 130 s after 10 s of digital silence, so in the aligned output the first sample above -60 dBFS after the silence must fall on the same frame as in the input.

6WHAT IS MEASURED

Everything graded is measured from the audio alone, so no access to the device's internals is needed. The same instrument measures the device's output and both reference processings.

• Meters: ITU-R BS.1770-4 K-weighting and gating, with the -70 LUFS absolute gate and the -10 LU relative gate.
• Clock: one reading every 100 ms of aligned output.
• Windows: momentary (400 ms), short-term (3 s), and gated integrated loudness over the most recent 3, 10, 30, 60 and 300 s (W3 to W300), over 900 s (WL), and over the whole case.
• Settling exclusion: the first 60 s of a case, and the 60 s after a reset, are not counted. WL is fed from the end of the exclusion, so it never contains the start-up.
• A counted reading is outside the exclusion, in a window holding at least one gating block that passes both gates. A window that silence has emptied is not counted.
• In band means within T ± 1 LU. Each window reports the share of counted readings inside, above and below the band, and the share not counted.
• Excursions: for each window, the longest run of counted readings outside the band, its length and its largest deviation; and the largest short-term reading above T.
• True peak: the BS.1770-4 true peak of the output, 4x oversampled, the maximum over the case.
• Noise floor: the 10th percentile of momentary loudness, digital silence left out, for input and output. Their difference is the pause lift: how far the adjuster raised the background.

If the device reports its gain

These figures need the applied gain and are reported where it is available:

• Gain movement, the sum of every change in dB per minute, and reversals, changes of direction of 0.5 dB or more: the objective stand-ins for "not heard working".
• Settling against the lag interval. At every change (the start, a join between parts, a reset) the gain may lag between the need before and the need after. The figure is how far the gain went outside that interval in the first 60 s, on each side. At the start and after a reset the need before is 0 dB. A pause's need is anything from the gain held before it to 0 dB.

7THE REFERENCE PROCESSINGS

A figure on its own says little: a percentage in band depends as much on the material as on the adjuster. So every case is also measured twice without an adjuster, by the same instrument.

• Unprocessed: the input as it is. The distance to it is what the adjuster adds.
• Fixed gain: the whole case scaled by one gain, T minus its integrated loudness, true-peak limited to the same ceiling. This is the best a single gain can do with the whole programme known in advance. The distance to it is what the content allows.
• Fixed gain per part: on synthetic and joined cases, one such gain for each part. It defines each part's need.

8THE CRITERIA

Five criteria are graded pass or fail. Everything else in section 6 is reported beside them, against the reference processings, without a pass line.

9WHAT A REPORT STATES

• The declarations of section 1.
• For each case and each criterion that applies: the measured value, the tolerance, and the verdict.
• The reported figures beside the same figures for both reference processings.
• Every item of programme material with its source, licence, length, sample peak and codec history, and a checksum of the copy that was measured.
• Every case left out or re-levelled, with the reason.

10WHAT THE METHOD DOES NOT COVER

• Formats beyond 44.1 and 48 kHz and two channels. Other rates and multichannel layouts need cases of their own.
• Fault input: DC, full-scale square waves, clipped audio, non-finite samples.
• Agreement between platforms or builds. A report states the one it measured.
• Dialogue-gated targets. Every level here is full-programme loudness.
• Audibility. Gain movement and reversals stand in for "not heard working". They are measurements, not a listening test.