Loudness accuracy: method & results

Meter Core v0.7.7's integrated-loudness readings (LUFS-I) were cross-checked against FFmpeg's EBU R128 loudness measurement, using a scripted, reproducible test set. This page is the receipt: every command, file hash and per-case number behind the claim.

Result: across 8 cases at 44.1, 48 and 96 kHz, the worst-case disagreement was 0.05 LU (mean 0.03 LU).

Reference & commands

Reference readings were taken with ffmpeg version 8.1.1. FFmpeg's loudnorm and ebur128 filters are implementations of EBU R128 loudness measurement (not an official reference implementation). Two independent FFmpeg paths were run per file and agreed within 0.07 LU on every case:

ffmpeg -hide_banner -nostats -i FILE -af loudnorm=print_format=json -f null -   # primary
ffmpeg -hide_banner -nostats -i FILE -filter_complex ebur128=peak=true -f null - # corroboration

In plain English: each audio file was measured twice by FFmpeg — a free, widely trusted audio tool used across the industry — using two of its separate loudness analyzers. Those readings became the benchmark; Meter Core then measured the same files, and the table below shows how closely the numbers agree. Anyone with FFmpeg installed can paste these commands and reproduce the benchmark themselves.

Test set

Two full mixes plus six synthetic signals designed so a wrong implementation cannot pass: calibrated 997 Hz stereo anchor tones whose integrated loudness is derived from ITU-R BS.1770's own calibration statement (near-exact by construction; FFmpeg read every anchor within 0.05 LU of its target), K-curve tones at 100 Hz and 10 kHz that verify the weighting-filter shape across the spectrum, and a loud/quiet/silence program that exercises both BS.1770 gates.

Meter Core capture protocol

Logic Pro, Meter Core v0.7.7 alone on an otherwise-empty stereo output, project sample rate matched to each file, meters reset per pass, one full uninterrupted playback per file, values taken from the plugin's CSV analysis export. Buffer-size independence was verified by re-running the −23 LUFS anchor at I/O buffers of 32 and 1024 samples: identical readings (−23.03 both).

Per-case results

Case	Rate	Length	SHA-256 (first 12)	FFmpeg I (LUFS)	Meter Core I	Δ (LU)
Full mix — "GULLY" master (D&B)	48 kHz	191 s	`79e7fcaa20f2`	-11.44	-11.39	+0.05
Full mix — AutoFactory promo	48 kHz	67 s	`22243def6504`	-12.27	-12.22	+0.05
Anchor tone 997 Hz @ −23 LUFS target	48 kHz	30 s	`59312d87b425`	-23.05	-23.06	-0.01
Anchor tone 997 Hz @ −14 LUFS target	44.1 kHz	30 s	`3b3a5e8f4e12`	-14.05	-14.05	+0.00
Anchor tone 997 Hz @ −23 LUFS target	96 kHz	30 s	`ee4baa9a59c0`	-23.05	-23.04	+0.01
K-curve tone 100 Hz	48 kHz	30 s	`59751a6d8dd6`	-21.85	-21.87	-0.02
K-curve tone 10 kHz	48 kHz	30 s	`6dad4527da31`	-16.65	-16.69	-0.04
Gating program (loud/quiet/silence/loud)	48 kHz	60 s	`f0224bb68eda`	-13.18	-13.22	-0.04

Worst-case |Δ| = 0.05 LU · mean |Δ| = 0.03 LU.

Scope — what this does and doesn't cover

Covered: integrated loudness (LUFS-I) as displayed live and in Track Pass / CSV export, Meter Core v0.7.7, stereo program at 44.1/48/96 kHz.
Not covered: short-term/momentary loudness, loudness range (LRA), surround layouts — not part of this cross-check.
True peak: Meter Core currently uses a lightweight, non-oversampled true-peak estimator; this page makes no accuracy claim for true peak. A standards-based oversampled true-peak stage (ITU-R BS.1770-5 Annex 2) is planned, and this page will be updated with its results when it ships.

Full raw artifacts (per-case JSON, capture CSVs) are archived internally with the hashes above; if you'd like the complete data set to check the working, email hello@autofactory.ai.

← Meter Core — free download