LEVELAPI2

Realtime adjuster results

THE REALTIME ADJUSTER,
MEASURED CASE BY CASE

LevelAPI2 202610170, Medium preset, in each of its three modes (50 ms, 2 ms, and 2 ms holding the dialogue), measured by the published test method on 117 cases at -23 LUFS and 68 programme cases at -16 LUFS: synthetic level jumps and pauses, bass tones, broadcast recordings, films, music, read speech, conversation and a live stream: 25 hours of material. Every graded line is shown with every case that fails it and the measured figure.

68/68Programmes on target, -23 LUFS, mode 0
45/45Dialogue on target, -23 LUFS, mode 2
1812/1815Block-size runs identical
50 / 2 msLatency, by mode
117 + 68Cases

WHAT WAS DECLARED

DeclarationValue
DeviceLevelAPI2 realtime adjuster, revision 202610170, preset Medium, every other setting at its default, in each mode of RealtimeAdjuster_SetMode(): 0, standard; 1, low latency; 2, low latency with dialogue
Target-23 LUFS; the programme cases again at -16 LUFS
True-peak ceiling-1 dBTP, the default (SetTruePeakCeilingDb)
Latency50 ms in mode 0, 2 ms in modes 1 and 2, measured on the audio: mode 0: 2400 frames at 48 kHz, 2205 at 44.1 kHz; mode 1: 96 frames at 48 kHz, 88 at 44.1 kHz; mode 2: 96 frames at 48 kHz, 88 at 44.1 kHz
Dialogue anchorMode 2: the calibration holds the dialogue's loudness rather than the programme's, both read by the library's speech detector. Below 10 % speech it holds the programme, above 25 % the dialogue, and blends between
Maximum boost+12 dB
Reference block10 ms (480 frames at 48 kHz, 441 at 44.1 kHz)
Block sizes declared identicalevery size, in every mode. Tested at 20 ms and 256, 512, 1024 and 2048 frames on every case but P6 and the conversation cases (108 cases in mode 0, 112 in modes 1 and 2, with the tones), and at 64 frames on the synthetic cases and the EBU / ITU files (49 in mode 0, 53 in modes 1 and 2)
Program lengththe whole stream by default; the long cases again with 900 s

GRADED, AT -23 LUFS

The method's pass lines, a column per mode. For mode 2, which holds the dialogue at the target, G1-D takes the place of G1 and G2, which are reported beside it; modes 0 and 1 report G1-D. A failing row stays on this page with its figure and what it is.

Pass lineScopeMode 0, standardMode 1, low latencyMode 2, low latency with dialogue
G1Whole-case integrated output within ±1 LU of ‑23 LUFS, every programme case68 of 68; 15 of 15 with a 900 s program length; the same at every block size68 of 68; 15 of 15 with a 900 s program length; the same at every block sizereported: 61 of 68; 15 of 15 with a 900 s program length; the same at every block size
G1-DOutput dialogue loudness within ±1 LU of the target, on the programme cases whose input is at least 25 % speech, both read by the library's speech detectorreported: 43 of 45reported: 44 of 4545 of 45
G2Long-term loudness in band on every case longer than 16 min13 of 15; 14 of 15 with a 900 s program length12 of 15; 13 of 15 with a 900 s program lengthreported: 10 of 15; 9 of 15 with a 900 s program length
G3Maximum true peak at or under ‑0.8 dBTP on the lossless cases (the synthetic signals, the tones and the EBU / ITU files), at every block size51 of 53; the highest passing case reads ‑0.99 dBTP53 of 53; the highest passing case reads ‑1.45 dBTP53 of 53; the highest passing case reads ‑1.59 dBTP
G4The last 30 s of each constant synthetic part within ±1 LU, where the need is within the +12 dB maximum67 of 76, the same parts at every block size67 of 7668 of 76
G5The output at every other block size identical to 10 ms blocks, sample for sample588 of 589 runs at 6 other sizes612 of 613 runs at 6 other sizes612 of 613 runs at 6 other sizes

Every failing row

LineCaseModeMaterialMeasuredWhat it is
G2P120, 1Live stream 1, game audio and voice0: 87.03 % in band
1: 88.10 % in band
A sparse live stream: over its quiet stretches (-35 to -58 LUFS) the gain sits on the +12 dB maximum and the output stays low. The fixed-gain two-pass reference reads 69.1 % here.
G2P13.31Conversation 3, panel discussion, TIB AV-Portal (CC BY 3.0 DE)91.63 % in bandAn 84-minute panel discussion recorded about 6 LU under the target. For about ten minutes the gain sits on the +12 dB maximum and the long-term loudness on the lower edge of the band. The 2 ms peak stages take about 0.15 dB more off this speech than the 50 ms mode, and on the maximum the calibration cannot give it back.
G2P60, 1Long run, four mixes back to back (workout, mixed, pop, rock)0: 92.24 % in band
1: 89.36 % in band
Three hours and 42 minutes of music with no program length: after two hours a new minute is too small a share of the long-term correction to move it. With 900 s it passes.
G3S6.10pink T-15, 10 s silence, bursts at T10 ms ‑0.46, 20 ms ‑0.46, 64 frames ‑0.46, 256 frames ‑0.46, 512 frames ‑0.46, 1024 frames ‑0.46, 2048 frames ‑0.46 dBTPWhite-noise bursts after a quiet passage and digital silence. The limiter holds -1 dBTP on its own 4x estimate; a high-precision true-peak meter can read white-noise bursts up to about 1.7 dB higher (0.54 dB on this case). The ceiling setting is there for this: set it that much lower for a guarantee on such material.
G3S6.20pink T-15, 10 s room tone at -50, bursts at T10 ms ‑0.46, 20 ms ‑0.46, 64 frames ‑0.46, 256 frames ‑0.46, 512 frames ‑0.46, 1024 frames ‑0.46, 2048 frames ‑0.46 dBTPAs S6.1, with room tone at -50 LUFS instead of silence.
G4S3.up part 20, 1P5, -12 dB -> 0 dB at 180 s0: ‑24.89 LUFS
1: ‑24.08 LUFS
Programme material split at its file joins, 30 to 83 s after a 12 dB step: the ride has not settled yet.
G4S3.up part 30, 1, 2P5, -12 dB -> 0 dB at 180 s0: ‑20.50 LUFS (the fixed-gain reference reads ‑20.50)
1: ‑20.62 LUFS (the fixed-gain reference reads ‑20.50)
2: ‑19.03 LUFS (the fixed-gain reference reads ‑20.50)
Programme material split at its file joins, 30 to 83 s after a 12 dB step: the ride has not settled yet.
G4S3.up part 40, 1, 2P5, -12 dB -> 0 dB at 180 s0: ‑24.80 LUFS
1: ‑24.48 LUFS
2: ‑21.93 LUFS
Programme material split at its file joins, 30 to 83 s after a 12 dB step: the ride has not settled yet.
G4S3.down part 10, 1, 2P5, 0 dB -> -12 dB at 180 s0: ‑21.25 LUFS
1: ‑21.04 LUFS
2: ‑21.62 LUFS
As S3.up, after a 12 dB step down.
G4S3.down part 20, 1, 2P5, 0 dB -> -12 dB at 180 s0: ‑25.20 LUFS
1: ‑25.13 LUFS
2: ‑24.51 LUFS
As S3.up, after a 12 dB step down.
G4S5.-60.30 part 30, 1, 2pink T-10, 30 s room tone at -60, pink T-100: ‑24.13 LUFS
1: ‑24.13 LUFS
2: ‑24.13 LUFS
Room tone in a pause of 30 s or more is lifted up to 12 dB. The returning programme meets that boost and is over-corrected afterwards, so the part ends up to 1.6 LU low.
G4S5.-60.120 part 30, 1, 2pink T-10, 120 s room tone at -60, pink T-100: ‑24.56 LUFS
1: ‑24.56 LUFS
2: ‑24.56 LUFS
Room tone in a pause of 30 s or more is lifted up to 12 dB. The returning programme meets that boost and is over-corrected afterwards, so the part ends up to 1.6 LU low.
G4S5.-50.120 part 30, 1, 2pink T-10, 120 s room tone at -50, pink T-100: ‑24.26 LUFS
1: ‑24.26 LUFS
2: ‑24.26 LUFS
Room tone in a pause of 30 s or more is lifted up to 12 dB. The returning programme meets that boost and is over-corrected afterwards, so the part ends up to 1.6 LU low.
G4S5.-45.120 part 30, 1, 2pink T-10, 120 s room tone at -45, pink T-100: ‑24.07 LUFS
1: ‑24.07 LUFS
2: ‑24.07 LUFS
Room tone in a pause of 30 s or more is lifted up to 12 dB. The returning programme meets that boost and is over-corrected afterwards, so the part ends up to 1.6 LU low.
G5S7 at 2048 frames0pink T-10, reset at 120 s85 of 3000 ticks differ from 120.0 sThe test harness resets the stream on the first host block boundary at or after 120 s. At 2048 frames that boundary falls 1024 frames later than at every other size, so the stream is reset at a different sample. Identical again from 140.8 s.
G5S7 at 2048 frames1, 2pink T-10, reset at 120 s1: 81 of 3000 ticks differ from 120.0 s
2: 81 of 3000 ticks differ from 120.0 s
The test harness resets the stream on the first host block boundary at or after 120 s. At 2048 frames that boundary falls 1024 frames later than at every other size, so the stream is reset at a different sample.

Outside a reported line

Where a mode's figure is outside a line that is reported for it rather than graded.

LineCaseModeMaterialMeasuredWhat it is
G1P4.12ITU BS.1770-2 conformance, stereo‑20.75 LUFS (dialogue ‑22.95)An ITU conformance piece whose dialogue sits under the mix: mode 2 holds the dialogue at the target, so the programme sits above it.
G1P4.22ITU BS.1770-2 conformance, mono‑21.12 LUFS (dialogue ‑23.16)An ITU conformance piece whose dialogue sits under the mix: mode 2 holds the dialogue at the target, so the programme sits above it.
G1P9.012Agent 327 (Blender, CC-BY)‑21.49 LUFS (dialogue ‑22.34)A film with little speech, where mode 2 blends the programme and the dialogue; the dialogue sits under the mix, so the programme lands above the target.
G1P9.082Elephants Dream (Blender, CC-BY)‑21.41 LUFS (dialogue ‑22.97)A film with little speech, where mode 2 blends the programme and the dialogue; the dialogue sits under the mix, so the programme lands above the target.
G1P9.142Sprite Fright (Blender, CC-BY)‑21.94 LUFS (dialogue ‑22.85)A film whose dialogue sits under its mix: mode 2 holds the dialogue at the target, so the programme sits above it.
G1P9.162The Daily Dweebs (Blender, CC-BY)‑21.67 LUFS (dialogue ‑22.53)A film whose dialogue sits under its mix: mode 2 holds the dialogue at the target, so the programme sits above it.
G1P122Live stream 1, game audio and voice‑24.17 LUFS (dialogue ‑22.54)A game stream whose voice sits well above the game audio: mode 2 holds the voice at the target, so the programme sits under it.
G1-DP9.160The Daily Dweebs (Blender, CC-BY)dialogue ‑24.55 LUFS (programme ‑22.48)A film whose dialogue sits under its mix: with the programme at the target, the dialogue sits under it.
G1-DP120, 1Live stream 1, game audio and voice0: dialogue ‑20.98 LUFS (programme ‑22.87)
1: dialogue ‑21.64 LUFS (programme ‑22.96)
A game stream whose voice sits well above the game audio: with the programme at the target, the voice sits above it.
G2P10.32Music mix 3, pop92.45 % in bandA music mix in which the speech detector hears speech. Mode 2 then holds that speech at the target; it sits under the mix, so the music is lifted, for tens of minutes. Mode 1 is the mode for music.
G2P10.52Music mix 5, workout57.65 % in bandA music mix in which the speech detector hears speech. Mode 2 then holds that speech at the target; it sits under the mix, so the music is lifted, for tens of minutes. Mode 1 is the mode for music.
G2P122Live stream 1, game audio and voice23.86 % in bandA game stream whose voice sits well above the game audio: mode 2 holds the voice at the target, so the programme sits under it.
G2P13.32Conversation 3, panel discussion, TIB AV-Portal (CC BY 3.0 DE)93.77 % in bandAn 84-minute panel discussion recorded about 6 LU under the target. For about ten minutes the gain sits on the +12 dB maximum and the long-term loudness on the lower edge of the band. The 2 ms peak stages take about 0.15 dB more off this speech than the 50 ms mode, and on the maximum the calibration cannot give it back.
G2P62Long run, four mixes back to back (workout, mixed, pop, rock)84.31 % in bandThe 3 h 42 min feed opens with the music of P10.5, which mode 2 lifts; with no program length the long-term loudness carries that start for hours.

WHAT A LISTENER HEARS, AND WHAT CHANGES IT

Where the adjuster is reliable, and what it does outside that range, with the modes each row concerns. "What changes it" names only what the shipped API offers, and what a host controls anyway.

Three excerpts of real programme, to hear in both 2 ms modes against the original, are on the listening examples page.

What the listener hearsCase that shows itModesWhat changes it
After a jump to louder material (an ad after a quiet programme) it is too loud for about half a minute, by up to 7.7 LU, then settlesS1 and S2 steps up: back in band in 26-48 s (S2.20, which arrives on the +12 dB maximum: 85 s), in mode 0; mode 1 is never louder over 3 sallNot adjustable. A host that knows the boundary (a playlist change, an ad insertion) can call Reset() there: the stream starts fresh from 0 dB (S7: back in band 9.8 s after a reset). Reset() allocates, so call it off the audio thread
After a drop to quieter material it stays quiet for up to a minute and a halfS1 and S2 steps down: back in band in 47-83 sallNot adjustable; Reset() at a known boundary as above
After a long loud passage (an explosion in a film), the quieter programme that follows can sit more than 6 LU under the target for tens of seconds. Modes 1 and 2 remove 63 % and 52 % of itEvery case at ‑23 LUFS, depth-weighted: 762 / 283 / 368 LU-seconds in modes 0 / 1 / 2; the longest, P9.12, 61 s in mode 0. Counted where the input is audible, the output more than 6 LU under the target and 3 LU under the fixed-gain reference, after a passage at least 8 LU louder in the last 30 sallModes 1 and 2
Material more than 12 LU under the target stays under itS1.-20 (never back in band), P12's quiet stretchesallNot adjustable: +12 dB is the most the adjuster adds
At a raised target, quiet material stays under it: at -16 LUFS, anything quieter than about -28 LUFS-16 LUFS run: P7.02 (‑17.65 LUFS), P13.3 (‑18.79 LUFS)allSetTargetLUFS(): a lower target leaves more material within +12 dB. The same cases pass at -23 LUFS
A quiet talk on the +12 dB maximum sits a little lower in the 2 ms modesP13.3, a panel discussion: long-term loudness in band 91.6 % in mode 1, 96.8 % in mode 01, 2A lower target, as in the row above
Background noise in a pause is raised (room tone, hiss, an audience), by up to 12 dB when the pause lasts 30 s or more. The programme then returns about 2 LU loud for 25-51 s, then up to 1.6 LU quietS5, room tone for 30 s or 120 s. Five seconds of room tone costs nothingallNot adjustable
After digital silence: pauses of 10 s or less return unchanged. After 30 s, material that needed a lift returns up to 4 LU quiet for about 6 s; after a minute or more the correction has faded back to 0 dB, and quiet material returns up to 10 LU under the target, back in band in about 12 sS4, silence of 2 to 120 s, in the 50 ms mode. Material at the target returns clean after every lengthallNot adjustable. A long pause is treated as the end of the programme
A stream switched on before its programme starts raises the room tone by 12 dB; the programme arrives up to 7.4 LU loud, then dips, back in band in 24-141 sS8, room tone for 30 to 300 s before the programme, in the 50 ms modeallStart the stream, or call Reset(), when the programme starts
A percussive entry after a quiet passage is up to 3.8 LU loud for about 18 sS6.1, S6.2, in the 50 ms modeallNot adjustable
Peaks read above the ceiling on a high-precision true-peak meter: up to about 1.7 dB on noise-like material, up to about 0.35 dB on lossy programme at -16 LUFS. In the 2 ms modes the limiter's estimate also covers the point between two samples, and the lossless cases stay under ‑0.8 dBTPS6: ‑0.46 dBTP in mode 0 at the default -1 dBTP, ‑1.59 in modes 1 and 2. At -16 LUFS: P13.4, ‑0.66 dBTPallSetTruePeakCeilingDb(): set the ceiling that much lower. Lossless programme material stays under -0.8 dBTP at the default
At a raised target, a bass line made of hard kicks lands under it in the 2 ms modes: their short limiter holds every kickL2.kick, a 50 Hz kick loop, at -16 LUFS: ‑17.32 LUFS in modes 1 and 2, ‑15.67 in mode 0; the fixed-gain two-pass reference reaches ‑17.42. At -23 LUFS: ‑22.23 / ‑23.00 / ‑23.00 in modes 0 / 1 / 21, 2Mode 0 for very peaky bass at a high target, or a lower target
In mode 2, music in which the speech detector hears speech is lifted: its 5-minute loudness sits more than 1 LU above the target for 24 min, up to 2.2 LUP10.5, a 1 h 00 min music mix (mode 1: 27 s above)2Mode 1 for music
In mode 2, the first minute after a start or Reset() settles less evenly: at some point its gain moves up to 6.0 dB further than mode 1's on speech. Over the first minute it sits as close to the target as the 50 ms mode: 3.3 LU from it on average, against 3.3 in mode 0 and 3.0 in mode 1. On the synthetic signals the difference from mode 1 is up to 3.5 dB, for at most 20 sThe 47 speech cases (at least 25 % speech), 10-60 s after the input is first audible2Not adjustable; mode 1 settles more evenly
On a feed that runs for hours, the level drifts up to about 1.3 LU after the first two hoursP6: long-term loudness in band 92.2 %, mode 0allSetProgramLengthSeconds(900): 99.9 % on the same feed; neutral on an hour of material
Peaks above the ceiling with the limiter offnot measuredallSetLimiterEnabled(true), the default

AGAINST THE REFERENCE PROCESSINGS

Median share of the time the output's loudness sits within ±1 LU of the target, over 10 s and 300 s windows, for modes 0 / 1 / 2 / a two-pass normalisation that knows the whole file in advance / the untouched input; and the median gain movement for modes 0 / 1 / 2. The two-pass figure is what a file-based workflow achieves; a live adjuster cannot know the future, and trades that for following the programme as it changes.

MaterialCasesW10 in band %W300 in band %Gain movement dB/min
Broadcast recordings1177 / 77 / 68 / 66 / 2100 / 100 / 100 / 100 / 04.0 / 4.0 / 4.2
Zapping stream162 / 62 / 62 / 25 / 2998 / 98 / 97 / 30 / 223.6 / 3.6 / 3.6
EBU / ITU reference programme424 / 38 / 9 / 11 / 13100 / 100 / 38 / 60 / 5915.1 / 15.4 / 13.8
Films1723 / 26 / 24 / 19 / 394 / 100 / 81 / 63 / 017.0 / 15.6 / 15.6
Music mixes558 / 58 / 49 / 53 / 095 / 93 / 89 / 94 / 02.5 / 2.5 / 2.6
Read speech, long360 / 60 / 61 / 51 / 3100 / 100 / 100 / 91 / 05.4 / 5.3 / 5.4
Poetry readings, short2060 / 62 / 59 / 62 / 0100 / 100 / 100 / 100 / 05.6 / 7.2 / 4.4
Conversation458 / 59 / 59 / 54 / 1100 / 100 / 100 / 77 / 05.0 / 5.2 / 5.2
Monologue over a bed155 / 54 / 54 / 37 / 399 / 98 / 94 / 57 / 14.4 / 4.6 / 4.7
Live stream123 / 29 / 23 / 22 / 1991 / 88 / 46 / 69 / 3120.6 / 22.2 / 17.9
Long run, 3 h 42 min156 / 55 / 54 / 31 / 185 / 84 / 75 / 37 / 02.8 / 2.8 / 2.7

THE SYNTHETIC CASES

Pink noise at defined levels, programme material stepped by 12 dB, silence, room tone and noise bursts. "Need" is the gain a perfect fixed gain would apply after the change; "over" and "under" are the largest excess of the short-term loudness over the minute after it. Each figure is mode 0 / mode 1.

CaseSignalChange at sNeed dBOver LUUnder LUAbove band sBack in band s
S1.+6pink T -> T+6180‑6.06.0 / 6.00.0 / 0.027.0 / 26.935.2 / 35.1
S1.+12pink T-2 -> T+10 (re-based)180‑10.07.3 / 5.50.1 / 0.042.1 / 44.342.6 / 45.2
S1.+20pink T-10 -> T+10 (re-based)180‑10.07.7 / 6.00.5 / 0.344.9 / 54.047.9 / 56.1
S1.-6pink T -> T-6180+6.00.0 / 0.05.4 / 5.40.0 / 0.050.2 / 50.2
S1.-12pink T -> T-12180+12.00.0 / 0.011.8 / 11.80.0 / 0.071.1 / 71.2
S1.-20pink T -> T-20180+20.00.0 / 0.020.0 / 20.00.0 / 0.0not in the case / not in the case
S2.6pink T-6 -> T180+0.05.5 / 5.50.4 / 0.424.8 / 24.826.2 / 26.2
S2.12pink T-12 -> T180+0.07.1 / 5.70.7 / 0.538.4 / 40.439.2 / 42.7
S2.20pink T-20 -> T180+0.07.2 / 3.87.1 / 7.318.2 / 15.084.9 / 30.2
S2.-6pink T+6 -> T180+0.00.0 / 0.05.3 / 5.30.0 / 0.047.0 / 47.0
S2.-12pink T+10 -> T-2 (re-based)180+2.00.4 / 0.411.4 / 11.40.0 / 0.068.3 / 70.1
S2.-20pink T+10 -> T-10 (re-based)180+10.00.4 / 0.420.4 / 20.40.0 / 0.083.4 / 83.4
S3.upP5, -12 dB -> 0 dB at 180 s180‑1.64.8 / 3.02.5 / 2.632.3 / 34.1not in the case / 49.8
S3.upP5, -12 dB -> 0 dB at 180 s246+0.03.9 / 3.92.2 / 1.73.7 / 3.8not in the case / not in the case
S3.upP5, -12 dB -> 0 dB at 180 s330+2.90.0 / 0.04.8 / 4.90.0 / 0.0not in the case / not in the case
S3.downP5, 0 dB -> -12 dB at 180 s180+10.41.7 / 2.48.3 / 7.84.5 / 5.5not in the case / not in the case
S3.downP5, 0 dB -> -12 dB at 180 s246+12.00.0 / 0.01.1 / 1.20.4 / 0.4not in the case / not in the case
S3.downP5, 0 dB -> -12 dB at 180 s330+14.80.0 / 0.04.8 / 4.41.0 / 0.0not in the case / not in the case
S4.T.2pink T, 2 s silence, pink T122+0.00.0 / 0.00.0 / 0.00.0 / 0.00.1 / 0.1
S4.T.10pink T, 10 s silence, pink T130+0.00.0 / 0.00.0 / 0.00.0 / 0.00.8 / 0.8
S4.T.30pink T, 30 s silence, pink T150+0.00.0 / 0.00.0 / 0.00.0 / 0.00.8 / 0.8
S4.T.60pink T, 60 s silence, pink T180+0.00.0 / 0.00.0 / 0.00.0 / 0.00.8 / 0.8
S4.T.120pink T, 120 s silence, pink T240+0.00.5 / 0.00.0 / 0.00.0 / 0.00.8 / 0.8
S4.L.2pink T-10, 2 s silence, pink T-10122+10.00.0 / 0.00.5 / 0.50.0 / 0.00.1 / 0.1
S4.L.10pink T-10, 10 s silence, pink T-10130+10.00.0 / 0.00.5 / 0.50.0 / 0.00.8 / 0.8
S4.L.30pink T-10, 30 s silence, pink T-10150+10.00.5 / 0.54.0 / 4.00.0 / 0.06.3 / 6.3
S4.L.60pink T-10, 60 s silence, pink T-10180+10.01.0 / 1.010.0 / 10.00.0 / 0.011.9 / 11.9
S4.L.120pink T-10, 120 s silence, pink T-10240+10.01.0 / 1.010.0 / 10.00.0 / 0.011.9 / 11.9
S5.-60.5pink T-10, 5 s room tone at -60, pink T-10125+10.00.0 / 0.00.5 / 0.50.0 / 0.00.1 / 0.1
S5.-60.30pink T-10, 30 s room tone at -60, pink T-10150+10.02.0 / 2.00.1 / 0.127.7 / 27.911.1 / 11.2
S5.-60.120pink T-10, 120 s room tone at -60, pink T-10240+10.02.0 / 2.01.0 / 1.050.8 / 50.9not in the case / not in the case
S5.-50.5pink T-10, 5 s room tone at -50, pink T-10125+10.00.0 / 0.00.5 / 0.50.0 / 0.00.1 / 0.1
S5.-50.30pink T-10, 30 s room tone at -50, pink T-10150+10.02.0 / 2.00.5 / 0.529.2 / 29.238.7 / 38.7
S5.-50.120pink T-10, 120 s room tone at -50, pink T-10240+10.02.0 / 2.01.0 / 1.040.9 / 40.942.9 / 42.9
S5.-45.5pink T-10, 5 s room tone at -45, pink T-10125+10.00.0 / 0.00.5 / 0.50.0 / 0.05.7 / 5.8
S5.-45.30pink T-10, 30 s room tone at -45, pink T-10150+10.02.0 / 2.00.3 / 0.325.1 / 24.931.3 / 31.1
S5.-45.120pink T-10, 120 s room tone at -45, pink T-10240+10.02.0 / 2.00.8 / 0.835.8 / 35.842.0 / 42.0
S6.1pink T-15, 10 s silence, bursts at T130‑0.03.8 / 2.31.8 / 0.917.9 / 17.628.7 / 17.7
S6.2pink T-15, 10 s room tone at -50, bursts at T130‑0.03.8 / 2.32.8 / 1.816.9 / 16.231.5 / 28.5
S8.-60.3030 s room tone at -60, then pink T30+0.07.4 / 4.44.8 / 1.723.5 / 26.623.6 / 91.6
S8.-60.120120 s room tone at -60, then pink T120+0.07.4 / 5.06.0 / 3.838.6 / 38.7141.1 / 98.2
S8.-60.300300 s room tone at -60, then pink T300+0.07.4 / 5.06.0 / 3.838.6 / 38.7141.1 / 98.2
S8.-50.3030 s room tone at -50, then pink T30+0.07.4 / 4.14.3 / 3.918.1 / 22.288.1 / 37.4
S8.-50.120120 s room tone at -50, then pink T120+0.07.4 / 5.45.9 / 3.923.6 / 23.824.1 / 24.5
S8.-50.300300 s room tone at -50, then pink T300+0.07.4 / 5.45.9 / 3.923.6 / 23.824.1 / 24.5
S8.-45.3030 s room tone at -45, then pink T30+0.07.4 / 4.14.3 / 3.016.7 / 14.182.2 / 26.9
S8.-45.120120 s room tone at -45, then pink T120+0.07.4 / 4.35.5 / 3.019.8 / 17.790.1 / 45.4
S8.-45.300300 s room tone at -45, then pink T300+0.07.4 / 4.35.5 / 3.019.8 / 17.790.0 / 44.9

Eight of them, traced

Top: output short-term loudness, with the ±1 LU band. Bottom: the gain applied, against the gain a perfect fixed gain would apply (dashed). Green: mode 0. Amber: mode 1. White: mode 2, drawn where it differs from mode 1 by more than 1 dB.

S1.+12 A 12 LU step up: too loud for about 40 s, then settled.
-13 -23 -33 +12 0 dB -12 0 s 60 s 120 s 180 s 240 s 300 s 360 s
S1.-12 A 12 LU step down: the lift arrives over about a minute.
-13 -23 -33 +12 0 dB -12 0 s 60 s 120 s 180 s 240 s 300 s 360 s
S2.20 Loud material after a long quiet stretch: the gain arrives on the +12 dB maximum. Mode 2 differs from mode 1 here by up to 4.0 dB.
-13 -23 -33 +12 0 dB -12 0 s 60 s 120 s 180 s 240 s 300 s 360 s
S3.up The same 12 dB step on programme material. Mode 2 differs from mode 1 here by up to 4.5 dB.
-13 -23 -33 +12 0 dB -12 0 s 60 s 120 s 180 s 240 s 300 s 360 s
S4.L.30 30 s of digital silence in material that needs a lift.
-13 -23 -33 +12 0 dB -12 0 s 60 s 120 s 180 s 240 s
S5.-50.120 Two minutes of room tone at -50 LUFS in a pause.
-13 -23 -33 +12 0 dB -12 0 s 60 s 120 s 180 s 240 s 300 s 360 s
S6.1 Noise bursts after a quiet passage and 10 s of silence. Mode 2 differs from mode 1 here by up to 2.5 dB.
-13 -23 -33 +12 0 dB -12 0 s 60 s 120 s 180 s
S8.-50.120 The stream starts on two minutes of room tone. Mode 2 differs from mode 1 here by up to 3.0 dB.
-13 -23 -33 +12 0 dB -12 0 s 60 s 120 s 180 s 240 s 300 s 360 s

BASS AND THE 2 MS LIMITER

Sine tones at 40, 63 and 100 Hz that step from 20 LU under the target to 10 LU over it, where the limiter takes the step, and a 50 Hz kick loop. Low-frequency distortion as the method defines it: harmonics 2 to 10 against the fundamental, Hann window. Reported, without a pass line.

CaseSignalTHD % 20-200 ms after the step, modes 0 / 1 / 2THD % 0.3-1.3 s after, modes 0 / 1 / 2Loudest momentary over the target LU, modes 0 / 1 / 2
L1.4040 Hz tone T-20 -> T+100.355 / 0.219 / 0.219 (input 0.051)0.408 / 0.001 / 0.0016.21 / 6.79 / 6.79
L1.6363 Hz tone T-20 -> T+100.070 / 0.183 / 0.183 (input 0.019)0.547 / 0.003 / 0.0037.71 / 7.91 / 7.91
L1.100100 Hz tone T-20 -> T+100.337 / 0.098 / 0.098 (input 0.000)0.492 / 0.001 / 0.0018.59 / 8.50 / 8.50

The kick loop

Integrated output, LUFS. The highest true peak at -16 LUFS is ‑1.00 dBTP: the limiter holds every kick at the ceiling.

TargetMode 0Mode 1Mode 2Fixed-gain two-pass reference
‑23 LUFS‑22.23‑23.00‑23.00‑23.00
‑16 LUFS‑15.67‑17.32‑17.32‑17.42

BLOCK SIZES

Every decision is taken on a 10 ms grid counted from the start of the stream, so the output does not depend on how the host cuts the stream into calls. Checked by a 64-bit hash of every output sample, per 100 ms. The last row of each mode is the library against the tuned lab chain it was ported from.

ModeBlock sizeCases identicalDifferences
020 ms108 of 108every output sample identical
064 frames49 of 49every output sample identical
0256 frames108 of 108every output sample identical
0512 frames108 of 108every output sample identical
01024 frames108 of 108every output sample identical
02048 frames107 of 108S7: from 120.0 s, 85 of 3000 ticks
0the library against the tuned lab chain, 10 ms45 of 45every output sample identical
120 ms112 of 112every output sample identical
164 frames53 of 53every output sample identical
1256 frames112 of 112every output sample identical
1512 frames112 of 112every output sample identical
11024 frames112 of 112every output sample identical
12048 frames111 of 112S7: from 120.0 s, 81 of 3000 ticks
1the library against the tuned lab chain, 10 ms117 of 117every output sample identical
220 ms112 of 112every output sample identical
264 frames53 of 53every output sample identical
2256 frames112 of 112every output sample identical
2512 frames112 of 112every output sample identical
21024 frames112 of 112every output sample identical
22048 frames111 of 112S7: from 120.0 s, 81 of 3000 ticks
2the library against the tuned lab chain, 10 ms117 of 117every output sample identical

CALL TIME AND MINIMUM BLOCK SIZE

The call that starts each 10 ms step does the adjuster's periodic work, so it is the slowest. The smallest host blocks at which that call takes at most half of the block, in 999 calls of 1000, single-threaded on an otherwise idle machine, which leaves the host the other half. Measured over 1 h 00 min of wide-ranging music (P10.1); the mean CPU on P9.15.

ModeLatencyMinimum block at 48 kHz, up to an hour after Create or ResetContinuous operation (estimated)Mean CPU per call, share of the block: 64 frames / 10 ms
0, standard50 ms64 frames (1.3 ms)128 frames4.2 % / 3.6 %
1, low latency2 ms64 frames (1.3 ms)128 frames4.2 % / 3.6 %
2, low latency with dialogue2 ms128 frames (2.7 ms)256 frames5.0 % / 4.3 %

Frames at 48 kHz; at 44.1 kHz the same frame counts leave more time. The adjuster's long-term meters cost more as the range of loudness they have heard since Create or Reset() grows, so a stream that runs for many hours without a reset needs the larger block. That column is estimated from the measured cost per meter step, not measured over days.

The step-start call, 99.9th percentile as a share of the block

Mode64 frames128 frames256 frames
0, standard48 %23 %14 %
1, low latency42 %22 %12 %
2, low latency with dialogue72 %38 %20 %

PROGRAM LENGTH ON LONG FEEDS

The cases longer than 16 minutes, with the default (the whole stream) and with a 900 s program length. Neutral on an hour of material; on three hours and more it keeps the long-term level on target.

CaseMaterialLengthLong-term in band %, mode 0: default / 900 sLong-term in band %, mode 1: default / 900 sLong-term in band %, mode 2: default / 900 sIntegrated LUFS at 900 s, modes 0 / 1 / 2W300 in band % at 900 s, modes 0 / 1 / 2
P2aZapping stream, recordings 1-1154 min 58 s99.1 / 99.199.5 / 99.599.5 / 99.5‑22.82 / ‑22.82 / ‑22.6999.5 / 99.5 / 96.5
P7Poetry, twenty readings in twelve languages40 min 23 s99.5 / 99.599.5 / 99.599.7 / 99.7‑23.07 / ‑23.08 / ‑23.02100.0 / 100.0 / 100.0
P8.2Audiobook, The Cat of Maurice39 min 50 s99.4 / 99.499.3 / 99.398.8 / 98.8‑23.17 / ‑23.19 / ‑23.06100.0 / 100.0 / 100.0
P10.1Music mix 1, classical1 h 00 min98.8 / 98.899.7 / 99.799.7 / 99.7‑22.79 / ‑22.80 / ‑22.8090.9 / 88.3 / 88.3
P10.2Music mix 2, chart pop1 h 03 min99.7 / 99.7100.0 / 100.0100.0 / 100.0‑22.50 / ‑22.49 / ‑22.4994.2 / 91.9 / 91.9
P10.3Music mix 3, pop1 h 10 min100.0 / 100.099.8 / 99.892.5 / 92.5‑22.39 / ‑22.39 / ‑22.3894.1 / 94.2 / 92.0
P10.4Music mix 4, rock and pop31 min 34 s100.0 / 100.0100.0 / 100.0100.0 / 100.0‑22.48 / ‑22.48 / ‑22.48100.0 / 100.0 / 100.0
P10.5Music mix 5, workout1 h 00 min99.5 / 99.599.6 / 99.657.6 / 63.0‑22.29 / ‑22.29 / ‑22.0599.6 / 99.6 / 58.6
P11Mixed 1, monologue over an intermittent bed59 min 33 s97.5 / 97.599.0 / 99.099.0 / 90.7‑22.75 / ‑22.77 / ‑22.6798.7 / 97.5 / 73.1
P12Live stream 1, game audio and voice30 min 36 s87.0 / 87.088.1 / 88.123.9 / 43.6‑22.89 / ‑23.00 / ‑23.5691.0 / 87.4 / 53.5
P13.1Conversation 1, European Commission midday briefing (CC BY 4.0)49 min 48 s99.9 / 99.999.9 / 99.9100.0 / 100.0‑23.10 / ‑23.12 / ‑23.09100.0 / 100.0 / 100.0
P13.2Conversation 2, FOMC press conference (public domain)28 min 42 s99.9 / 99.999.9 / 99.999.9 / 99.9‑23.12 / ‑23.11 / ‑23.10100.0 / 100.0 / 100.0
P13.3Conversation 3, panel discussion, TIB AV-Portal (CC BY 3.0 DE)1 h 24 min96.8 / 96.791.6 / 91.493.8 / 93.8‑23.43 / ‑23.53 / ‑23.4785.8 / 80.5 / 84.2
P13.4Conversation 4, US Supreme Court oral argument (public domain)1 h 49 min99.8 / 99.899.9 / 99.999.9 / 99.9‑22.68 / ‑22.71 / ‑22.68100.0 / 100.0 / 100.0
P6Long run, four mixes back to back (workout, mixed, pop, rock)3 h 42 min92.2 / 99.989.4 / 99.984.3 / 90.1‑22.45 / ‑22.45 / ‑22.3694.5 / 94.9 / 76.9
P6 Long-term loudness over 3 h 42 min with a 900 s program length: mode 0, mode 1, mode 2; dashed, mode 0 with the default. The ±1 LU band.
-21 -23 -25 0 min 30 min 60 min 90 min 120 min 150 min 180 min 210 min

AT -16 LUFS

The programme cases and the kick loop again at the streaming target. The other synthetic cases stay at -23 LUFS, where their levels are defined.

Pass lineScopeMode 0, standardMode 1, low latencyMode 2, low latency with dialogue
G1Whole-case integrated output within ±1 LU of ‑16 LUFS63 of 6763 of 67reported: 59 of 67
G1-DOutput dialogue loudness within ±1 LU, cases with at least 25 % speechreported: 42 of 45reported: 43 of 4543 of 45
G2Long-term loudness in band, cases longer than 16 min11 of 1411 of 14reported: 10 of 14
G3Maximum true peak at or under ‑0.8 dBTP on the lossless cases run at this target (the EBU / ITU files and the kick loop)5 of 5; highest ‑0.88 dBTP5 of 5; highest ‑1.00 dBTP5 of 5; highest ‑0.97 dBTP

Every failing row

LineCaseModeMaterialMeasuredWhat it is
G1P7.020, 1Poetry, French 20: ‑17.65 LUFS
1: ‑17.49 LUFS
Needs +12.4 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.
G1P7.030, 1Poetry, German 10: ‑17.28 LUFS
1: ‑17.26 LUFS
A 132 s reading that needs a large lift: it ends before the ride has settled.
G1P9.020, 1Big Buck Bunny (Blender, CC-BY)0: ‑17.39 LUFS
1: ‑17.32 LUFS
A quiet film with loud peaks: the true-peak ceiling holds it back, not the boost maximum.
G1P13.30, 1Conversation 3, panel discussion, TIB AV-Portal (CC BY 3.0 DE)0: ‑18.79 LUFS
1: ‑18.76 LUFS
Needs +13.0 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.
G1-DP7.022Poetry, French 2dialogue ‑17.48 LUFS (programme ‑17.50)Needs +12.4 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.
G1-DP13.32Conversation 3, panel discussion, TIB AV-Portal (CC BY 3.0 DE)dialogue ‑18.88 LUFS (programme ‑18.71)Needs +13.0 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.
G2P8.20, 1Audiobook, The Cat of Maurice0: 89.73 % in band
1: 90.92 % in band
Needs +11.2 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.
G2P120, 1Live stream 1, game audio and voice0: 50.16 % in band
1: 68.31 % in band
The sparse live stream again, further under a higher target.
G2P13.30, 1Conversation 3, panel discussion, TIB AV-Portal (CC BY 3.0 DE)0: 0.11 % in band
1: 0.13 % in band
Needs +13.0 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.

Outside a reported line

LineCaseModeMaterialMeasuredWhat it is
G1P4.12ITU BS.1770-2 conformance, stereo‑14.37 LUFS (dialogue ‑16.17)An ITU conformance piece whose dialogue sits under the mix: mode 2 holds the dialogue at the target, so the programme sits above it.
G1P4.22ITU BS.1770-2 conformance, mono‑14.54 LUFS (dialogue ‑16.16)An ITU conformance piece whose dialogue sits under the mix: mode 2 holds the dialogue at the target, so the programme sits above it.
G1P7.022Poetry, French 2‑17.50 LUFS (dialogue ‑17.48)Needs +12.4 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.
G1P7.032Poetry, German 1‑17.34 LUFS (dialogue ‑16.97)A 132 s reading that needs a large lift: it ends before the ride has settled.
G1P9.022Big Buck Bunny (Blender, CC-BY)‑17.32 LUFS (dialogue ‑24.22)A quiet film with loud peaks: the true-peak ceiling holds it back, not the boost maximum.
G1P9.082Elephants Dream (Blender, CC-BY)‑14.70 LUFS (dialogue ‑16.39)A film with little speech, where mode 2 blends the programme and the dialogue; the dialogue sits under the mix, so the programme lands above the target.
G1P122Live stream 1, game audio and voice‑17.42 LUFS (dialogue ‑15.80)A game stream whose voice sits well above the game audio: mode 2 holds the voice at the target, so the programme sits under it.
G1P13.32Conversation 3, panel discussion, TIB AV-Portal (CC BY 3.0 DE)‑18.71 LUFS (dialogue ‑18.88)Needs +13.0 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.
G1-DP7.020, 1Poetry, French 20: dialogue ‑17.59 LUFS (programme ‑17.65)
1: dialogue ‑17.44 LUFS (programme ‑17.49)
Needs +12.4 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.
G1-DP9.160The Daily Dweebs (Blender, CC-BY)dialogue ‑17.27 LUFS (programme ‑15.72)A film whose dialogue sits under its mix: with the programme at the target, the dialogue sits under it.
G1-DP13.30, 1Conversation 3, panel discussion, TIB AV-Portal (CC BY 3.0 DE)0: dialogue ‑18.96 LUFS (programme ‑18.79)
1: dialogue ‑18.92 LUFS (programme ‑18.76)
Needs +13.0 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.
G2P8.22Audiobook, The Cat of Maurice91.48 % in bandNeeds +11.2 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.
G2P10.52Music mix 5, workout71.36 % in bandA music mix in which the speech detector hears speech. Mode 2 then holds that speech at the target; it sits under the mix, so the music is lifted, for tens of minutes. Mode 1 is the mode for music.
G2P122Live stream 1, game audio and voice14.60 % in bandA game stream whose voice sits well above the game audio: mode 2 holds the voice at the target, so the programme sits under it.
G2P13.32Conversation 3, panel discussion, TIB AV-Portal (CC BY 3.0 DE)0.15 % in bandNeeds +13.0 dB. Near or beyond the +12 dB the adjuster adds at most, so the target is reached slowly or not at all.

THE MATERIAL

Everything but the synthetic signals, the tones and the EBU / ITU files came through lossy sources, so it is kept out of the true-peak grading (G3): a codec's own inter-sample overshoot cannot be told apart from the device's.

HOW IT WAS RUN