Why two loudness meters show different LUFS.

Two different tools share the name "loudness meter"

Search for a loudness meter and you get two unrelated products in one list. One reads a digital file and returns LUFS, the perceived-loudness scale streaming platforms use. The other listens through a microphone and returns dB SPL, the loudness of the air in your room. They are not the same measurement, and a microphone reading can never predict a file's LUFS:

• Playback gain. Turn your monitors up 6 dB and the room reading moves 6 dB. The file didn't change.
• The room. Walls, corners and your listening position add and cancel energy that was never in the audio.
• Uncalibrated microphones. A phone capsule has its own frequency response and no reference level. Two phones on the same table disagree.
• Clipping. Push a phone microphone hard enough and it distorts, so the reading flattens out.

File meters, on the other hand, do agree. Feed one file to two meters that both implement the ITU-R BS.1770 loudness algorithm and the integrated LUFS should match to within rounding. Disagreement means someone measured a different span of the file, used a different setting, or isn't following the standard. It doesn't mean loudness is a matter of opinion.

Every conforming meter reads the same file the same

BS.1770 spells out the K-weighting filter, the 400 ms block size, the gates and the averaging. Feed the same whole file to two meters that follow it and the arithmetic has one answer. That is why "which loudness meter is most accurate" is a badly formed question: a meter either implements the standard or it doesn't. So don't trust anyone on this, including us. The test takes about ten minutes:

1. 01Pick two files: a full-length track, and a short file that starts with a few seconds of silence. The silent intro exposes any meter that reports ungated loudness.
2. 02Drop each into the online analyzer and write down integrated LUFS, true peak and LRA.
3. 03Run the same files through the loudness meter in your DAW, and through ffmpeg with the command below.
4. 04Compare the integrated figures. Select the entire file every time, or you are measuring different audio and the test proves nothing.

ffmpeg -nostats -i yourfile.wav -filter:a ebur128=peak=true -f null -

If the integrated numbers line up, every meter is doing the maths correctly and the question is closed for good. True peak is the one reading where implementations can differ a little, because meters oversample and filter the signal in different ways. Expect small gaps there, a few hundredths of a dB. Expect none on integrated loudness.

So why did your two meters disagree? The four real causes

Four setup differences explain nearly every gap. The numbers below are an example: a three-minute stereo master that reads −9.8 LUFS integrated, measured wrong on purpose.

• A selection, not the whole file. Highlight just the last chorus and the reading jumps to around −7.9 LUFS. Nothing broke: you measured twelve seconds of the loudest part. Clear the selection before you compare anything.
• Gating switched off. If the track opens with a quiet intro, an ungated reading drags the number down, to −11.4 in our example. BS.1770 uses two gates: an absolute gate at −70 LUFS that drops digital silence, and a relative gate 10 LU below the running mean that drops the quiet passages. A tool without both gates answers a different question.
• A different window. Momentary loudness averages 400 ms and short-term averages 3 seconds. The loudest momentary value can read −6.1 while the integrated value is still −9.8. Comparing your momentary peak with someone else's integrated number is the most common version of this mistake.
• A different channel count. Measure only the left channel of a stereo pair and you throw away half the summed power. 10 × log10(2) = 3.01, so similar material reads about 3 LU quieter: −12.8 instead of −9.8.

Work the list in order: measure the whole file in both tools, confirm both show integrated rather than momentary, check both see the same channel layout, and make sure you are on the same bounce. If all four match and the readings still differ by more than a tenth of a decibel, you have found a genuinely non-conforming meter, which is much rarer than the four causes above.

The same trap catches mono files. BS.1770 adds the channels together rather than averaging them, so a mono podcast and its stereo-duplicated export do not measure the same: the stereo version reads about 3 LU louder. Measure the file in the exact channel layout you will publish.

Why your intro, fade and podcast pauses move the number

Integrated loudness doesn't average your whole file. It throws part of it away first. The relative gate isn't a fixed level: it sits 10 LU below your programme's own loudness. A podcast around −18 LUFS gates at about −28 LUFS, and anything quieter isn't counted at all.

Take an example episode (assumed values, to show the mechanism): speech at −18 LUFS, a 15-second ambient intro at −35, room-tone pauses at −75 and a slow fade at the end.

• The intro selected on its own reads about −35 LUFS. The gate compares the intro with itself, so nothing is dropped.
• Intro plus two minutes of speech reads roughly −18. The intro sits 17 LU below the mean, so the gate removes it.
• The whole episode still reads roughly −18. The pauses fall under the absolute gate and the fade tail drops out too.
• The bounced file with silence at head and tail reads the same again. That silence costs you nothing.

Quiet material either survives the gate or vanishes. The trouble is right at the edge: a quiet bridge at −27 LUFS in a −18 programme sits 9 LU down, so it counts. At −29 it's gone. Trim your selection a second earlier and the number moves for no visible reason. The fix is boring and it works: measure the exact file you will upload, start to finish.

Where the meter goes, and readings that don't count

Anything you insert after the meter changes the file but not the reading. Put a limiter after the meter and the number you are watching describes audio nobody will hear. Pre-limiter and post-limiter readings differ by everything the limiter does: if it adds 4 dB of loudness, a pre-limiter meter reads about 4 LU quieter than your export. Put the meter last.

Real-time meters have a second trap. Integrated loudness is a running total, so it describes the file only once the whole file has played through. Stop at 2:10 of a 3:30 song and you have a reading for 2:10 of audio, and the meter won't tell you it's incomplete.

The export dialog gets the last word, too. Normalize on export rescales the file, so a carefully hit target is gone before upload. Dither adds very low-level noise when you drop to 16-bit, so the export isn't bit-identical to what your session meter saw. Make the exported file the thing you measure: bounce, then drop the bounce into a file analyzer. A file measurement can't be truncated, can't sit before a limiter and can't miss a checkbox.

One more for live metering on a Mac: system sounds and other apps count if they play through the same output. In LOUDMETER: Menu Bar, Ignore Apps lets you choose apps the meter should not hear, such as a video call or your mail app, while everything else on your monitor output is measured normally.

Reading the numbers, and what each one changes

• Integrated is the whole file as one number. It decides how much gain a platform adds or takes away. A file reading −17 LUFS against a −14 target is 3 dB short: one fader move, not a remix.
• Short-term and momentary are the same measurement over 3 seconds and 400 ms. Use them to find the loudest few seconds, then listen there: that is where your limiter works hardest and distortion shows first.
• Loudness range (LRA) shows how far the quiet parts sit below the loud ones. There is no universally correct LRA, so compare with a reference you trust, measured yourself.
• True peak is the only number that predicts real damage. Keep it at −1 dBTP: lossy encoders shift sample values, so a file sitting at 0.0 can clip after encoding.
• PSR and PLR show what the limiter cost you. Push 3 dB more into a limiter with a fixed ceiling and the peak stays put while short-term loudness climbs about 3 dB, so PSR drops by about 3. That isn't a rule of thumb, it's subtraction.

Platforms publish different targets, and they behave differently too. YouTube turns loud files down and never turns quiet ones up. Spotify offers Loud, Normal and Quiet modes, and listeners can switch normalization off. The platform targets table lists the published values and dates them, because specs change.

Should you master louder than the target?

On a normalized path, extra loudness is gone. Take two masters of the same mix: A reads −8 LUFS, B reads −14, and the platform plays at −14. A is turned down 6 dB, B isn't touched, and both reach the listener at the same level. What A paid for that is roughly 6 dB of limited peaks, and turning the file down doesn't bring them back. Same volume, fewer transients.

There are paths where louder really is louder: DJ playback, radio, audio in a video you publish yourself, and listeners who switched volume-matching off. Nobody knows what share of listening runs through each path, so be suspicious of anyone who quotes one.

A simple rule: make one master at the target unless you can name the non-normalized path in one sentence before you touch the limiter. "A DJ is playing this off a USB stick" is a reason. "My competition is louder" isn't, because on streaming your competition is turned down too. If you can name the path, make a second, louder file for it and measure both.

When the gain you need would push true peak past −1 dBTP, three fixes usually cost less than more limiting: clean up the sub energy below your lowest real note, clip the one transient that sets the peak instead of limiting the whole bar, and turn down the single loud element in the mix that is setting the ceiling.

LUFS, LKFS and dialogue-gated loudness

LKFS stands for Loudness, K-weighted, relative to Full Scale. It's the same unit as LUFS under a different name: US broadcast paperwork says LKFS, European paperwork says LUFS, and one equals the other.

The real difference is the gate. A programme-gated reading measures everything that survives gating: speech, music, effects. A dialogue-gated reading measures only the parts a speech detector flags as talking, which is what Netflix specifies. On a talk-only podcast the two land close together; add loud music beds or an action mix and they can sit several LU apart. Loudmeter reports programme-gated loudness, so don't report its number against a spec that asks for the dialogue gate. When in doubt, ask the network or post house which gate and which spec revision they want.

Pick by the job: browser, session meter or the one in your DAW

• A browser analyzer. The Loudmeter online analyzer reads a WAV, MP3, AIFF, FLAC or M4A and reports integrated, short-term and momentary LUFS, true peak with the overs marked on a waveform, LRA, PSR, PLR and stereo correlation, and plays the file at the level Spotify, Apple Music, YouTube and other services will. Nothing to install, nothing uploaded. What it can't do is show a moving number while you mix.
• A meter inside your session. Loudness plug-ins show loudness moving as you move a fader. The cost is that you have to play the audio to measure it, one track at a time.
• The meter already in your DAW. Logic Pro ships a Loudness Meter that shows momentary, short-term and integrated loudness to EBU R128, and many other DAWs include loudness metering. Check your metering menu before you download anything.

File in hand, use a file analyzer. Fader in hand, use a meter in the session. If you have twelve finished podcast episodes to check, loading a session per file is the slow way to answer a question about a file.

A browser analyzer's real limit is your device's memory, because it decodes the whole file before measuring. A stereo 24-bit 48 kHz WAV is 288 kB per second, so ninety minutes is about 1.55 GB on disk and about 2 GB once decoded to 32-bit float. Use the browser for songs, stems and episodes, and a desktop app for long-form recordings.

For live metering there is LOUDMETER: Menu Bar for Mac ($9.99, one-time), which meters whatever your Mac is playing with no plug-in. The iPhone and iPad app is free: it analyzes files on the device and mirrors your Mac's live meters over your local network.

FAQ

The meter was never the variable. This week, take one finished master and one short file that opens with a few seconds of silence, measure both three ways (your DAW, a file analyzer and ffmpeg), and write the integrated figures down with the date next to them. Matching numbers mean your tools are fine, and every gap after that is a setting you can name.

Measure your file now.

Free, in your browser. Your file never leaves your device.