The Loudness War in My Own Library: Normalizing 4,800 Tracks Without Killing Dynamics

Last spring I finally ripped my entire CD collection to FLAC — 4,800 tracks, roughly 320 GB, spread across 210 albums. When I shuffled the whole library for the first time, the volume jumps were brutal: a 1988 pressing of a jazz trio sat at about -22 LUFS integrated, while a 2015 remaster of a rock album hit -8 LUFS. Every other track I was reaching for the volume knob. That’s not a music problem, that’s a signal-chain problem, and I wanted to fix it without flattening the quiet parts that make those old recordings breathe.

The tool I settled on was ffmpeg with the loudnorm filter, run in two passes so it actually measures before it adjusts. Pass one writes a log:

ffmpeg -i input.flac -af loudnorm=I=-16:TP=-1.5:LRA=11:print_format=json -f null -

That prints measured integrated loudness, true peak, and LRA. Pass two consumes those numbers:

ffmpeg -i input.flac -af loudnorm=I=-16:TP=-1.5:LRA=11:measured_I=-22.3:\
measured_TP=-0.8:measured_LRA=9.4:measured_thresh=-33.1:offset=0.0 \
-ar 44100 output.flac

I wrapped it in a small shell loop that checks whether the output already exists and skips re-encoding, so a crashed run at track 3,000 is resumable. I target -16 LUFS because that’s roughly where my DAC and headphones sound neutral; -1.5 dBTP leaves headroom for inter-sample peaks. LRA=11 keeps me from squashing anything with real dynamic range.

The catch is that loudnorm in single-pass mode is not what you want — it guesses, and on tracks with long quiet intros it guesses wrong. Two-pass is the only version I trust. Also, re-encoding lossy formats through this filter is a bad idea; I only run it on FLAC sources and keep the originals in a separate _raw/ tree, because once you’ve adjusted gain you can’t cleanly undo it without the source.

Where it breaks: anything already brickwalled at -9 LUFS gets pulled down, which is correct, but a track with a 20-second spoken intro and a loud chorus will still confuse the integrated measurement — the LRA guard helps but doesn’t eliminate it. I hit that on a live album where the applause between songs counted as loud content, and the quieter ballads got lifted too much. I ended up hand-marking those five tracks and leaving them alone.

What’s your threshold for “this needs normalizing” versus “this is the recording and I should leave it”? I’m curious whether people treat compilations differently from studio albums.