normy

Batch loudness normalization for MP3 collections, built for variable-quality archive material — old radio recordings, podcasts, audiobooks. Wraps FFmpeg's two-pass loudnorm filter (EBU R128) with a single-pass fallback, preserves source bitrates and sample rates per-file (so a 1 GB low-bitrate archive doesn't bloat into 30 GB), and resumes safely after Ctrl+C.

What it looks like

=== normy ===
Source:   /Volumes/Wolfpack/Media/Radio/Art Bell
Output:   /Volumes/Wolfpack/Media/Radio/Art Bell Normalized
Files:    921
Target:   -16 LUFS
Encoding: preserve source (floor 96k)
Started:  Sun Apr 26 16:52:45 EDT 2026
=============
[1/921] [ETA --] 1992-12-12 - Area 51 - John Lear - Bob Lazar.mp3
  OK ▲5.8dB (-21.79 → -16 LUFS, 96k @ 22050Hz) in 25s
[2/921] [✓1] [ETA 6h21m] 1993-06-20 - Al Bielek - Philadelphia Experiment.mp3
  OK ▲8.2dB (-24.20 → -16 LUFS, 64k → 96k @ 44100Hz) in 18s
[3/921] [✓2] [ETA 5h47m] 1993-09-03 - John Lear - UFOs.mp3
  OK ≈ (-16.30 → -16 LUFS, 96k @ 22050Hz) in 21s
[4/921] [✓3] [ETA 5h33m] 1993-10-30 - Ghost to Ghost 1993.mp3
  Encoding  ████████████░░░░░░░░  62% │ 01:12 / 01:56 │ 9.2x
...
[30/921] [✓27 ✗1 ⊘1] [ETA 4h12m] 1994-10-30 - Ghost To Ghost 1994.mp3
  OK ▼0.1dB (-16.12 → -16 LUFS, 96k @ 22050Hz) in 19s [! 87 decode warnings]
...

=== Done ===
Finished:  Sun Apr 26 22:14:09 EDT 2026
Elapsed:   5h21m
Total:     921
Success:   905
Skipped:   10
Failed:    6

Files with significant decode warnings (output may have audible glitches):
  - 1994-10-30 - Ghost To Ghost 1994.mp3 (87 warnings)
  - 1995-03-19 - Linda Moulton Howe.mp3 (102 warnings)

Bitrate distribution:
    32k: 47 files
    64k: 312 files
    96k: 245 files
   128k: 198 files
   192k: 87 files
   256k: 16 files

Requirements

• ffmpeg and ffprobe
• python3

brew install ffmpeg          # macOS
sudo apt install ffmpeg      # Debian/Ubuntu

Usage

./normy.sh                                  # interactive prompts
./normy.sh /path/to/mp3s                    # source only
./normy.sh /path/to/mp3s /path/to/output    # source + output

If you omit the output path, normy writes to a normalized/ subfolder inside the source. Last-used paths are remembered in ~/.normy_history and offered as defaults the next time you run it interactively.

Resuming

Press Ctrl+C anytime. normy finishes cleanup on the in-progress file, discards any partial output, and exits. Re-run with the same paths and it picks up where it left off. Press Ctrl+C twice to force quit.

Configuration

The defaults are tuned for spoken-word archive material. To change them, edit the constants near the top of normy.sh.

Setting	Default	Notes
Target loudness	-16 LUFS	Podcast / spoken-word standard
True peak ceiling	-1.5 dBTP
Loudness range	11 LU
Encoding mode	preserve	Match source bitrate, with floor
Bitrate floor	96 kbps	Bumps up ultra-low-bitrate sources
Sample rate	source	Preserved per-file from the input

Encoding modes

• preserve (default) — match each file's source bitrate, with the floor for very low ones. Best for archive material with mixed source quality; won't inflate small files.
• vbr — LAME VBR (-q:a 4 ≈ 165 kbps avg). Best for podcast/voice workflows where source quality is consistent.

Output

• Files mirror the source folder structure under the output directory.
• Each run writes a timestamped log to <output>/logs/normalize-YYYYMMDD-HHMMSS.log.
• <output>/normalize.log is a symlink to the current run's log, so tail -f normalize.log always tracks the active run.
• Existing output files are skipped, which is what makes Ctrl+C / rerun cheap.

License

The Unlicense — public domain. Do anything you want with this; no attribution required.