Media Ffmpeg Normalize

Media Ffmpeg Normalize

Context: $ARGUMENTS

ffmpeg-normalize is a Python wrapper around ffmpeg's loudnorm filter that automates proper 2-pass EBU R128 normalization, batch directories, and preserves non-audio streams. Raw loudnorm is 1-pass by default (dynamic, inaccurate); this tool does the measure-then-apply pass automatically.

Quick start

• Normalize one file (EBU R128 -23 LUFS default): ffmpeg-normalize in.mp3 -o out.mp3 → Step 3
• Streaming target (-14 LUFS YouTube/Spotify): ffmpeg-normalize in.wav -t -14 -o out.wav → Step 2
• Batch a podcast folder: ffmpeg-normalize episodes/*.mp3 -of out/ -t -16 -tp -1.5 -lrt 11 → Step 3
• Video (keep video stream, re-encode audio only): ffmpeg-normalize in.mp4 -c:a aac -b:a 192k -o out.mp4 → Step 3

When to use

• Batch-normalizing many files to one target loudness (podcast archive, music library, course videos).
• You want proper 2-pass EBU R128 without wiring the loudnorm JSON round-trip yourself (see ffmpeg-audio-filter skill if you need manual control).
• Delivering to broadcast, streaming platforms, or archives with an LUFS spec.
• Don't use for cross-file consistency (e.g., album-gain style). Each file is normalized independently. For consistent relative levels across a set, use a single manual 2-pass loudnorm with a shared measurement, per the ffmpeg-audio-filter skill.

Step 1 — Install

pip install ffmpeg-normalize
# or with uv
uv tool install ffmpeg-normalize

Requires ffmpeg on PATH. Verify:

ffmpeg-normalize --version
ffmpeg -hide_banner -filters | grep loudnorm

Step 2 — Pick a target

Decide the three EBU R128 parameters:

Rule of thumb: broadcast = -23 LUFS, podcast = -16 LUFS, streaming = -14 LUFS. See references/normalize.md for the full platform table.

Step 3 — Batch

Single file:

ffmpeg-normalize input.mp3 -o output.mp3 -t -16 -tp -1.5 -lrt 11

Folder (glob expanded by shell):

ffmpeg-normalize in/*.mp3 -of out/ -t -16 -tp -1.5 -lrt 11 --progress

Video (copy video, re-encode audio — REQUIRED since raw can't be muxed back):

ffmpeg-normalize in.mp4 -c:a aac -b:a 192k -o out.mp4

The helper script:

uv run ${CLAUDE_SKILL_DIR}/scripts/normalize.py preset --input in.mp3 --output out.mp3 --platform apple-podcasts
uv run ${CLAUDE_SKILL_DIR}/scripts/normalize.py batch --indir episodes/ --outdir normalized/ --platform spotify

Dry-run first:

ffmpeg-normalize in.mp3 -o out.mp3 --dry-run

Step 4 — Verify

Measure the output to confirm it hit the target:

ffmpeg -i out.mp3 -af ebur128=peak=true -f null - 2>&1 | tail -20
# Look for: I: (integrated LUFS), LRA, True peak

Or use --print-stats / --debug during the normalize run to emit per-file loudnorm JSON.

Available scripts

• scripts/normalize.py — check / normalize / preset / batch subcommands wrapping ffmpeg-normalize; stdlib only.

Reference docs

• Read references/normalize.md for the platform target table, LUFS/dBFS/LRA concepts, raw-loudnorm vs wrapper comparison, and recipe book.

Gotchas

• 2-pass is automatic, no JSON round-trip needed (unlike raw loudnorm).
• Default target is EBU R128 -23 LUFS (broadcast) — too quiet for streaming. Use -t -14 for YouTube/Spotify or -t -16 for Apple Podcasts.
• -t = integrated loudness (LUFS); -tp = true peak (dBFS, typically -1 to -2); -lrt = loudness range target (LU).
• Batch mode normalizes each file INDEPENDENTLY. There is no cross-file awareness. To keep relative levels across an album or episode set, do a single manual 2-pass loudnorm with a shared measurement (see ffmpeg-audio-filter).
• -c:a aac is REQUIRED for MP4 output (or pcm_s16le for WAV). Default encoder is pcm_s16le which only works in raw/WAV containers.
• Re-encodes audio by default — lossy for MP3/AAC inputs (transcode penalty). For FLAC/WAV output, no additional loss.
• -c:v copy is the default for video — the video stream is never re-encoded.
• -of output/ (output folder) creates the dir if missing. Naming convention: output/<original-filename>.<ext>.
• --keep-loudness-range-target preserves source LRA (don't compress dynamics).
• --dual-mono auto-detects dual-mono and compensates (-3 dB, per EBU spec).
• --lower-only only reduces level, never boosts — useful when you don't want to amplify quiet source noise.
• --extra-input-options / --extra-output-options pass arbitrary flags through to ffmpeg (e.g., -ar 44100).
• --progress shows a tqdm progress bar.
• Single-pass (dynamic) is the default for video audio to save time; force 2-pass with --keep-lra-above-loudness-range-target when source LRA exceeds target.
• Never chain a volume= filter after normalization — it defeats the normalization.
• LUFS = Loudness Units relative to Full Scale (perceptual loudness, K-weighted).
• dBFS = decibels Full Scale (absolute sample level).
• True Peak prevents inter-sample clipping on consumer DACs that upsample.
• LRA (Loudness Range) = statistical dynamic range in LU.

Examples

Example 1: Podcast library to Apple Podcasts spec

ffmpeg-normalize podcast/*.mp3 \
  -of podcast-normalized/ \
  -t -16 -tp -1.5 -lrt 11 \
  -c:a libmp3lame -b:a 192k \
  --progress

Example 2: Video for YouTube (keep video, AAC audio)

ffmpeg-normalize talk.mp4 \
  -t -14 -tp -1.0 \
  -c:a aac -b:a 192k \
  -o talk-yt.mp4

Example 3: Dry-run across a directory (plan only)

ffmpeg-normalize *.wav -of out/ -t -23 --dry-run --print-stats

Example 4: Only lower (never boost) for a noisy archive

ffmpeg-normalize noisy/*.wav -of safe/ -t -23 --lower-only

Example 5: Specific audio stream (e.g., second language track only)

ffmpeg-normalize in.mkv -o out.mkv --audio-stream-filter "a:1" -c:a aac

Troubleshooting

Error: Output file does not contain any stream or missing audio in MP4

Cause: Default codec pcm_s16le can't be muxed into MP4. Fix: Add -c:a aac -b:a 192k (or -c:a libmp3lame for MP3 output).

Output is quieter than expected

Cause: Default target is -23 LUFS (broadcast). Streaming targets are louder. Fix: Use -t -14 (YouTube/Spotify) or -t -16 (Apple Podcasts).

All files in a batch sound different relative to each other

Cause: Expected — each file is normalized to the target independently. Quiet originals get boosted to target; loud originals get cut. Fix: Use manual 2-pass loudnorm with a single concatenated measurement, or process all files together via the ffmpeg-audio-filter skill.

True peak still clips on consumer device

Cause: -tp default is -2 but inter-sample peaks can still exceed on some chains. Fix: Tighten to -tp -1.5 or -tp -2.0; do not go above -1.0.

RuntimeError: Could not find ffmpeg

Cause: ffmpeg not on PATH. Fix: Install ffmpeg (brew install ffmpeg on macOS) or set FFMPEG_PATH=/path/to/ffmpeg.