LUFS-S max

The LUFS-S max column follows LUFS-M max and reports the loudest complete 3-second window in a file. The numeric value has one decimal place; hovering gives the maximum window's start and end times. It describes a sustained local loudness maximum, complementary to whole-programme LUFS and the faster 400 ms momentary measurement.

Calculation#

The measurement uses the same K-weighted channel powers as integrated loudness and the same 3-second rolling history as LRA. For each decoded frame after the first three seconds:

S = −0.691 + 10 × log10(mean K-weighted channel-summed power over 3 s)

Following EBU Tech 3341, section 2.2, S uses a rectangular window with no gate or additional attack/release smoothing. The file maximum is evaluated at every complete sample-aligned window, rather than only when an LRA sample is collected. The largest power sum determines the location; exact ties keep the first.

The maximum is captured before the LRA calculation appends its analysis-only tail. There are no padded startup windows or measurements beyond the final real frame. The momentary loudness page describes the shared implementation, sample timing and column preferences.

Availability and limits#

• Supported audio is mono or stereo at 8–768 kHz, including decoded DSD PCM. Multichannel audio has no reading because reliable channel-position weights are not yet supplied.
• A file shorter than three seconds has no S maximum, even if it has valid integrated or momentary loudness. Exactly three seconds permits one complete S window.
• Silence, invalid input, failed/skipped decoding and older reports have no reading. Quiet nonzero audio below the integrated loudness gate remains measurable.
• The shared 3-second history stores powers as 32-bit floats to bound memory; its rolling sum uses compensated 64-bit arithmetic. Extreme finite float samples whose weighted power cannot fit invalidate S and LRA; M can remain available because its history uses 64-bit powers. Powers too tiny for the compact representation cannot produce a finite S reading.
• A maximum is not a typical loudness or a quality grade. Short accents are diluted over three seconds. A longer window can also average across pauses or transitions; audition the interval shown in the tooltip.
• K-weighting filter history affects boundaries. A maximum from an excerpt need not equal the maximum calculated over that same passage within the full file.
• This is not LRA. LRA describes the gated distribution of many S readings; this column retains the highest ungated S reading. Its value is in LUFS, whereas LRA is in LU.

Saved reports#

JSON stores loudness_peaks.short_term as { "lufs": number, "start_secs": number }, or null independently of the momentary reading. If neither reading is available, loudness_peaks is null. Older reports without this field remain readable.

CSV exports max_short_term_lufs and short_term_max_start_s after the momentary fields, before loudness_range_lu. Times are seconds from the first decoded frame. Missing readings stay blank; finite values retain their stored precision. The table uses the unrounded number for sorting and always puts unavailable readings last.

Verification#

cargo test -p flaccompagnon-core loudness_peaks
cargo test -p flaccompagnon-core --test loudness_reference -- --ignored --nocapture

The unit suite includes the alternating-level rectangular-window case 9 and all 20 shifted-file cases from EBU Tech 3341 case 10, the known stereo calibration level, the approximately 3.01 LU mono/stereo difference, audio below the integrated gate, exact three-second availability (including 8 and 11.025 kHz), recovery after a large finite float passage and invalid input. It also verifies that LRA's synthetic filter tail cannot increase exported M/S maxima. A separate WAV test checks a known level step, maximum positions, report fields and backward compatibility. The optional FFmpeg comparison covers mixed frequencies, several rates and both supported layouts.

Manual fixtures#

A sustained three-second burst#

ffmpeg -n -f lavfi -i 'aevalsrc=if(between(t\,2.15\,5.15)\,0.0707945784\,0)*sin(2*PI*1000*t)|if(between(t\,2.15\,5.15)\,0.0707945784\,0)*sin(2*PI*1000*t):s=48000:d=8' -c:a pcm_s24le short-term.wav
ffmpeg -hide_banner -nostats -i short-term.wav -af ebur128 -f null -

Expect LUFS-S max ≈ −23.0 with its window near 2.15–5.15 s. LUFS-M max should also be about −23.0. Small differences in timing are possible because of the K-filter response. In Audacity, a three-second stereo 1 kHz tone at linear amplitude 0.0707945784, with surrounding silence, produces the equivalent check.

Too short for an S reading#

ffmpeg -n -f lavfi -i 'aevalsrc=0.0707945784*sin(2*PI*1000*t)|0.0707945784*sin(2*PI*1000*t):s=48000:d=2' -c:a pcm_s24le too-short-for-s.wav

Expect integrated LUFS and LUFS-M max near −23.0, and a dash for LUFS-S max. To see how a brief accent is averaged differently, use the 400 ms burst on the momentary loudness page.