Loudness-normalize audio with FFmpeg (EBU R128, ATSC A/85)

When to use this

You loudness-normalize audio for broadcast delivery (EU broadcasters require EBU R128 compliance, NA broadcasters require ATSC A/85), multi-source content where different sources have different loudness (compilations, podcasts, news inserts), and any delivery context where viewers shouldn't reach for the volume knob between assets. Loudness normalization isn't the same as peak normalization — peak normalization equalizes the loudest sample, loudness normalization equalizes perceptual loudness across the whole asset using the LUFS standard. For broadcast you want loudness normalization.

Command variants

ffmpeg -i input.mp4 \
  -af "loudnorm=I=-23:LRA=7:TP=-2" \
  -c:v copy \
  output.mp4

ffmpeg -i input.mp4 \
  -af "loudnorm=I=-24:LRA=6:TP=-2" \
  -c:v copy \
  output.mp4

# Pass 1: measure
ffmpeg -i input.mp4 -af loudnorm=I=-23:LRA=7:TP=-2:print_format=json -f null -

# Output gives JSON with measured_I, measured_LRA, measured_TP, measured_thresh, target_offset.
# Use those values in pass 2:
ffmpeg -i input.mp4 \
  -af "loudnorm=I=-23:LRA=7:TP=-2:measured_I=-26.21:measured_LRA=8.5:measured_TP=-1.2:measured_thresh=-37.4:offset=0.05:linear=true" \
  -c:v copy \
  output.mp4

What each parameter does

• I=-23 / I=-24

  Integrated loudness target in LUFS. -23 LUFS is EBU R128 (Europe), -24 LUFS is ATSC A/85 (North America). YouTube/Netflix/Spotify use different targets (-14 LUFS).

• LRA=7

  Loudness Range. EBU R128 requires LRA ≤ 7 LU; ATSC A/85 requires LRA ≤ 6 LU. Limits the dynamic range to prevent fatigue.

• TP=-2

  True peak ceiling in dBTP. -2 dBTP prevents inter-sample clipping that decoders / DACs introduce. -1 dBTP is acceptable for some use cases.

• measured_*

  Pass-1 values fed into pass-2 for linear normalization. Required for two-pass mode.

• linear=true

  Linear normalization mode (vs dynamic). Required for two-pass; preserves the original dynamic range while shifting the integrated loudness.

• -c:v copy

  Stream-copy video. Loudness normalization is audio-only; never re-encode video as a side-effect.

What this outputs

Audio normalized to target LUFS with controlled true-peak. Single-pass deviates ~5-10% from target (good enough for streaming); two-pass deviates ~1% from target (required for broadcast contractual delivery). Video stream is unchanged.

Pitfalls

1. Single-pass loudnorm is convenient but inaccurate. EBU R128 broadcast contracts often specify "Integrated Loudness within ±0.5 LU of -23 LUFS" — single-pass typically misses this. Always two-pass for contractual delivery.
2. Wrong target for the platform: -23 LUFS for streaming feels too quiet (Spotify/YouTube/Netflix all use -14 LUFS). Match the target to the delivery platform, not the broadcast standard.
3. Dynamic mode (default) compresses audio toward the target, changing the perceived sound. Linear mode preserves the dynamic but shifts the level. Production delivery typically wants linear.
4. Inter-sample peaks vs sample peaks: TP measures inter-sample (true peak); standard peak meters measure sample peaks. -2 dBTP is more conservative than -2 dBFS sample peak.
5. Loudnorm processes the entire file in memory for two-pass mode. Very long audio (>4 hours) may need chunked processing.
6. Re-encoding video alongside audio normalization wastes compute. Always use -c:v copy.

At production scale

Audio normalization is fast at scale — typically 5-10% of the equivalent video re-encoding cost. Two-pass loudnorm doubles the audio processing time but is necessary for broadcast delivery. At >100K assets/month with broadcast-grade normalization, the audio processing cost is meaningful but still small relative to video compute. The pattern that scales: loudness-normalize once at the master encoding stage, stream-copy the audio to all subsequent renditions in the ABR ladder.

MpegFlow models loudness normalization as a parameterized DAG stage with the LUFS target as a per-job input. The two-pass measure-then-correct pattern is wrapped as a single stage that the platform manages internally. Audit log records the measured-I and target-I values so contractual delivery can be verified.

• FFmpeg
• audio
• loudness
• ebu-r128
• Broadcast
• audio-operations

• Recipe
  Extract Audio From Video
• Architecture
  Broadcast Grade Vod Transcoding