The measurements, and where they come from
Integrated loudness follows ITU-R BS.1770: the signal is K-weighted, split into 400-millisecond blocks at 75 per cent overlap, and gated twice — an absolute gate at −70 LUFS, then a relative gate ten units below the ungated mean. Short-term loudness uses the three-second window defined in EBU Tech 3341. Loudness range follows EBU Tech 3342, which uses a different relative gate: twenty units, not ten. That difference is a common implementation bug, and it makes loudness range read too narrow when it is got wrong.
True peak, and why oversampling is not optional
Sample values do not describe the waveform a converter reconstructs between them. Peaks that exist only between samples are real, and lossy encoders reconstruct them. ITU-R BS.1770 Annex 2 sets a four-times oversampling minimum for true-peak measurement; this tool uses eight. If your meter reports sample peak only, it is telling you about the file rather than about playback, and it will let a master through that distorts on every service.
PLR and PSR — useful, and not standards
Peak to loudness ratio is true peak minus integrated loudness across the whole track. Peak to short-term ratio is true peak measured against the loudest three-second window, which is where limiting actually bites. Neither is defined by ITU, EBU or AES. They are mastering conventions, and the thresholds people use — under five units meaning heavy limiting, under three meaning severe — come from practice, not from a document. This tool reports both and says plainly that they are conventions.
What playback normalisation really does
Every major service turns loud content down toward a target and generally does not turn quiet content up. The honest problem is that only some of those targets are published. Spotify documents its figure. EBU R 128 and the AES streaming recommendation are published standards. The numbers circulated for the other services are repeated everywhere and documented nowhere we could verify at source. The table on this page marks each row accordingly, because a confident number from an unverifiable source is worse than an honest gap.
The codec test nobody else runs
Lossy encoders do not preserve sample values. Quantisation and the decoder reconstruction filter mean the decoded waveform can peak above the source — routinely by half a decibel to a decibel and a half on dense material. A master sitting at full scale therefore clips after encoding even though the file never clipped. Most analysers guess at this. This one actually encodes your audio to AAC and Opus in your browser, decodes it back, and measures the decoded peak, so the number you see happened rather than being estimated.
Mono is not a formality
Correlation between the two channels tells you whether they reinforce or cancel. But the number that matters is the level you actually lose when the mix is summed, and this tool measures that directly rather than inferring it from correlation. Phone speakers, smart speakers, laptop playback and most club systems sum to mono or something close to it, so a wide mix that loses several decibels on summing is losing them in front of most of the audience.
What is a good PSR value?
There is no official answer, which is why the tool labels it a convention. In practice, a peak to short-term ratio above about eight units reads as a dynamic master, five to eight as controlled, three to five as heavily limited, and under three as severely limited. Whether that is wrong depends entirely on the music. The measurement tells you what you did; it does not tell you what you should have done.