loudnorm=I=-14:TP=-1.5:LRA=11 lifts a quiet voice track to about -14 LUFS in one pass, and aresample=48000 keeps it at 48 kHz instead of 96. ReelSDK runs the same filter for $0.02 per compute minute, under a fifth of a cent for this clip.
Normalize audio loudness with ffmpeg loudnorm
loudnorm measures loudness like a streaming player and sets the gain to hit a target: -14 LUFS and a -1 dB true peak for Spotify. ReelSDK's API got within a fifth of a LU in one pass; a measured second pass lands on it.
GuidesUpdated
ffmpeg -i in.mp4 -c:v copy \
-af loudnorm=I=-14:TP=-1.5:LRA=11,aresample=48000 \
-c:a aac -b:a 192k loud.mp4Good enough for most uploads. The video is copied, the audio is normalized and resampled back to 48 kHz, and the result lands within a fifth of a LU of the target.
Variations
The other commands you will need
Pass 1: measure
Nothing is written; loudnorm prints what it measured as JSON at the end of the log.
measure ffmpeg -i in.mp4 -af loudnorm=I=-14:TP=-1.5:LRA=11:print_format=json -f null -Pass 2: apply the measured values
Paste the input_* values back in with
linear=true. One gain for the whole file, so the dynamics stay as they were. This run printedNormalization Type: Linearand ended at -13.8 LUFS, -1.3 dBTP.second pass ffmpeg -i in.mp4 -c:v copy -af loudnorm=I=-14:TP=-1:LRA=11:\ measured_I=-26.21:measured_TP=-13.50:measured_LRA=2.20:\ measured_thresh=-36.79:offset=0.19:linear=true:print_format=summary,\ aresample=48000 -c:a aac -b:a 192k linear.mp4
The flags
The loudnorm options in these commands
| Flag | What it does | Used here |
|---|---|---|
I=-14 | Integrated loudness target in LUFS; the default is -24 | All three |
TP=-1.5 | Maximum true peak in dBTP; the default is -2 | All three |
LRA=11 | Loudness range target in LU; the default is 7 | All three |
print_format=json | Prints the measured and output values at the end | Pass 1 |
measured_I and the rest | The input values from pass 1; all four are needed for linear mode | Pass 2 |
offset | Gain applied before the true peak limiter; pass 1 prints it as target_offset | Pass 2 |
linear=true | One gain for the whole file instead of a moving one | Pass 2 |
aresample=48000 | Brings the sample rate back down after loudnorm | All writes |
Source: ffmpeg filters: loudnorm, checked November 2026.
One pass or two
In one pass loudnorm does not know what is coming, so it rides the gain as it goes. Its log calls this dynamic normalization. It is close but not exact: our clip landed at -14.2 LUFS against a -14 target.
Two passes fix that. The first run only measures. The second gets those numbers and, with linear=true, applies one fixed gain to the whole file, the way you would turn a volume knob once. Loud and quiet parts keep their distance from each other.
Why the linear pass came back dynamic
Our first try at pass 2 used TP=-1.5 and the log still said Normalization Type: Dynamic. The docs list the reason: linear mode is dropped when the gain would push the true peak above the target.
Our track needed about 12 dB of gain to reach -14, and that lifted its loudest peak to about -1.3 dBTP, above the -1.5 ceiling, so loudnorm switched modes. With the ceiling at TP=-1 the same command ran linear. The dynamic pass still landed on target, so both are usable; linear is the one that leaves the performance alone.
Which target to pick
Spotify plays everything at -14 dB LUFS and asks for masters below -1 dB true peak, and -14 is a common target for social video too. loudnorm's own defaults are quieter: -24 LUFS, a 7 LU range and -2 dBTP. Set I, TP and LRA yourself so the result does not depend on defaults.
LRA is the loudness range loudnorm aims for, and linear mode needs it to be at least the measured range. Speech has a narrow range (our clip measured about 2 LU), so 11 is never the limit.
Keep the sample rate
The docs say dynamic mode upsamples to 192 kHz to catch true peaks. Without a resampler our AAC file came out at 96 kHz, twice the rate of the source. aresample=48000 at the end of the chain, or -ar 48000, puts it back.
Without installing ffmpeg
The same command as one API call
Send the one-pass args to POST /v1/ffmpeg. Each job runs ffmpeg 8.1 on a fresh machine at $0.02 per compute minute, billed per second; this one finished in a few seconds, cost under a fifth of a cent and measured -14.2 LUFS. A finished job returns files, not ffmpeg's log, so run the measuring pass where you can read its output and send only the second pass.
curl "https://api.reelsdk.com/v1/ffmpeg?wait=60" \
-H "Authorization: Bearer $REELSDK_KEY" \
--json '{
"inputs": {
"in.mp4": "https://api.reelsdk.com/f/lib/in/guides/quiet-voice.mp4"
},
"args": [
"-i", "in.mp4", "-c:v", "copy", "-af",
"loudnorm=I=-14:TP=-1.5:LRA=11,aresample=48000", "-c:a", "aac", "-b:a",
"192k", "loud.mp4"
],
"outputs": ["loud.mp4"]
}'FAQ
Questions, answered
How do I normalize audio to -14 LUFS with ffmpeg?
Run ffmpeg -i in.mp4 -c:v copy -af loudnorm=I=-14:TP=-1.5:LRA=11,aresample=48000 -c:a aac out.mp4. For an exact result, measure first with print_format=json and pass the values back with linear=true.
Why does loudnorm say Dynamic when I set linear=true?
Linear mode needs all four measured values and enough headroom: if the gain would push the true peak above TP, or the target LRA is below the measured one, loudnorm falls back to dynamic. Raise TP a little or lower the target I.
Why is my output 96 kHz after loudnorm?
loudnorm upsamples internally to find true peaks. Add aresample=48000 after it, or -ar 48000, to write the rate you want.
Can I run loudnorm with an API?
Yes, POST the args to https://api.reelsdk.com/v1/ffmpeg. A job returns files rather than ffmpeg's log, so use one pass there, or measure locally and send only the second pass with the values filled in.
- Duck music under a voice with ffmpegLower background music only while someone speaks, with sidechaincompress keyed by the voice and amix to put them together.
- Mix two audio tracks with ffmpegLay music under a voice or combine any two tracks with amix, set their levels with weights, and avoid the three defaults that bite.
- Remove silence from audio with ffmpegTrim dead air at the start, shorten long pauses with silenceremove, or cut the matching video with select and aselect.
- Extract audio from a video with ffmpegCopy the AAC track out of an MP4 without re-encoding, or convert it to MP3 or WAV, with the file size each command produced.
- Replace the audio in a video with ffmpegSwap a video's soundtrack for another file with -map and -shortest, keep the picture untouched, and pad audio that runs short.
- Add music to videoMix a looped music track under a video's own audio with ducking and a fade-out in one ffmpeg call, for a fraction of a cent.
- FFmpeg onlineRun your own ffmpeg command on your files in the browser, free and without installing or uploading anything.
- ReelSDK vs self-hosted FFmpegTranscoding a minute of 1080p to 720p takes about 24 seconds of compute: about 0.8 cents on ReelSDK, about a tenth of a cent on a busy 4 vCPU box. Below about 11,000 minutes a month, the API costs less than an always-on box.
Sources, checked November 2026
Ready to build?
Ship video from your app today
Your first $0.50 is free, and one key covers every endpoint.
