A file can peak at 0 dBFS and still sound quiet
Peak is one sample; a plosive can hit full scale while RMS and integrated loudness stay low, which is the pairing the meter exists to show.
Peak is one sample; a plosive can hit full scale while RMS and integrated loudness stay low, which is the pairing the meter exists to show.
The Loudness Meter reports three figures from the file you drop: peak dBFS, whole-file RMS, and a BS.1770-style integrated estimate. Peak is the single loudest sample. 0 dBFS means that sample sat at full scale. It does not mean the programme is loud. A click, a mic pop, or one hard consonant can pin the peak while the rest of the take lives 20 dB down. If you only watch the peak, you will either crush a quiet read to "fix clipping" or ship a mix that sounds thin on a phone because you trusted the red light.
Peak is one sample
The analyser walks every channel, every sample, and keeps max(abs(x)). Then peakDbfs = 20 × log10(peak). One sample at 1.0 is 0 dBFS. One sample at 0.5 is about −6 dBFS. Everything else in the file is ignored by that number.
That is why a voiceover can "clip" in the meter and still sound small. The plosive on "pack" hits 0. The vowels sit at −18. RMS, which squares and averages every sample, stays low. Integrated loudness, which K-weights, blocks, and gates, stays low. Peak is what clips the converter. Loudness is closer to what you hear. The pairing is the product. A file can be both peak-limited and quiet.
The UI labels peak "Loudest sample" on purpose. It is not true peak. There is no oversampling. Inter-sample peaks after the DAC are not in this number. Do not use 0 dBFS here as a delivery stamp. Use it as "something in this file already occupies the ceiling."
RMS and integrated are the other two numbers
Average RMS is whole-file energy: 20 × log10(sqrt(sum of squares / count)). Silence in the file pulls it down. A 30-second read with 10 seconds of pad looks quieter than the same read trimmed. That is arithmetic, not a mix note. If you needed to cut the pad first, that is the audio trimmer — a cut list and a re-encode, not a loudness pass.
Integrated (est.) is the BS.1770-shaped figure: pre-filter and RLB biquads (K-weighting), 400 ms blocks, 100 ms hop, absolute gate near −70 LUFS, then a relative gate 10 dB below the ungated mean. The card also prints how many blocks survived gating, plus duration, sample rate, and channel count. Surround channels above the first two are weighted 1.41, matching the BS.1770 idea.
It is labelled an estimate. Filters run in JavaScript. decodeAudioData has already resampled, often to 48 kHz. There is no dialog-specific gate and no true-peak path. Broadcast delivery still wants Youlean, Nugen, or ffmpeg ebur128. This tab is the pairing: peak versus integrated, on the file you already have.
A limiter can park at 0 dBFS and still be quiet
Short-form exports often look "hot" because a limiter kissed 0 on the loudest sample and left the body of the mix alone. You open the meter. Peak: 0.0 dBFS. RMS: −18. Integrated: −20 LUFS or worse. The file is not loud. It is peaked. Turning the master up from there only recreates the clip. The fix is not another generate. The fix is the mix: take the plosive down, high-pass the pop, or ride the body up after the transient is handled, then measure again.
Loudness normalisation is the other half of this pairing on a platform: a hot peak can pull a whole programme down on ingest, and the quiet body gets quieter. A file that peaks at 0 and meters quiet is how that happens. Hitting 0 is not "podcast loud." Integrated is.
Do not prompt a video row to "make it −14." The model will invent a mix. It will not give you these three numbers. Drop the export. Read peak versus integrated. Then decide whether the job is a limiter, a trim, or a new take.
Measure the file you already have
Audio here is capped at 80 MB and 20 minutes once decoded. Compressed files expand into PCM in memory. A long podcast is heavy even when the mp3 looks small. That guard is local. Nothing is uploaded.
/cost is the hub for generates billed in credits. This meter is not on it. Credits buy models. Peak versus RMS is a read on a file.
If you have not generated yet, pick the video or speech row first, then meter the export. Paying for a 15-second miss to learn the last one was a plosive at 0 is how a measurement becomes a bill. If the take is the thing you are buying, generate. If the question is "why does this already-bounced VO feel small," open the meter.
Disconnect the network. The three numbers still appear.
FAQ
Why is my peak 0 dBFS when the file still sounds quiet?
Peak is the single loudest sample. A click or a plosive can hit 0 dBFS while RMS and integrated loudness stay low. Loudness is closer to perceived level; peak is what clips. That pairing is what the meter is for.
If peak is 0, should I turn the whole file down?
Not first. Find the sample that pinned the ceiling. If it is one consonant, fix that event. Turning the master down because one pop hit 0 makes the quiet body quieter. Measure integrated after the transient is handled.
Is the integrated number EBU R128?
No. It follows the BS.1770 idea — K-weighting, 400 ms blocks, absolute and relative gates — without true-peak oversampling, in JavaScript, after the decoder has resampled. Use a certified meter for a delivery spec.
Can I ask a model to hit a LUFS target instead of measuring?
No. A generate invents a mix. It does not report these three figures. Bounce the file you have. Read peak versus integrated here. Then mix or regenerate on purpose.