True peak, explained properly

This is the one technical concept in delivery where the standard explanation is either too vague to act on or too mathematical to read. It is neither difficult nor optional, and getting it wrong is the most common reason a good master arrives at the listener damaged.

Two different peaks

A digital audio file is a list of numbers. Each number is the amplitude of the signal at one instant, forty-four thousand or forty-eight thousand times a second. The sample peak is simply the largest of those numbers, and that is what most meters show you.

But nobody listens to a list of numbers. To play the file, a converter reconstructs a continuous waveform that passes through those points — and between any two samples, that curve can go higher than either of them.

The true peak is the highest point of the reconstructed waveform, including everything happening between the samples. It is the level that actually exists at the output. It is often higher than the sample peak, and on dense modern masters it can be higher by a decibel or more.

Sample peak is what you wrote down. True peak is what comes out of the speaker.

Where the extra level comes from

A small mental picture helps. Imagine a sine wave near full scale whose peak happens to fall exactly between two sample points. Both samples land slightly below the top of the curve, so the sample meter reports a level below the actual maximum. The reconstruction, which has to pass smoothly through both points, arcs up over the top — and that arc is a real signal, present in the analogue output.

The more high-frequency content and the sharper the transients, the more often peaks fall between samples and the larger the discrepancy. Which is why this problem grew with modern production: bright, dense, heavily limited masters generate inter-sample peaks constantly.

Why the ceiling is not zero

Zero dBFS is the largest number the format can represent. Anything above it cannot be stored, so it is clipped — but as we have just established, what is stored and what is reproduced are different things.

A file whose samples all sit at or just under zero can reconstruct to plus one, and every stage downstream that works on the reconstructed signal has to deal with that. Converters clip it. Lossy decoders clip it. Some processing stages clip it silently. None of them ask your permission.

Setting the ceiling at minus one dBTP leaves a decibel of room for the reconstruction to arc into. That is the whole reasoning. It is not superstition, it is not headroom for a mastering engineer to use later, and it is not wasted loudness — on a normalised platform the gain gets made up on playback anyway.

What it sounds like

Inter-sample clipping is easy to miss because it is brief and it does not sound like obvious distortion. It sounds like a hardness on loud transients, a faint crackle on sibilance, a sense that the top end is brittle on some systems and fine on others.

That last part is the trap. Different converters, different decoders and different phones handle overshoot differently, so the fault is present on some playback and absent on others. If a record sounds clean in your room and harsh on a friend's phone, this is one of the first things to check.

Measuring it

A true peak meter oversamples — reconstructs the waveform at four times the sample rate or more — and reports the peak of that reconstruction. Any compliant meter does this; the specification for it lives in the same document as LUFS loudness measurement, and it is standard across every serious tool.

Measure after every stage that could change the signal. In particular, measure after the final limiter and again after any sample rate conversion, because conversion changes the reconstruction and can move the true peak. A file that measured at minus one before a rate conversion is not guaranteed to measure at minus one after it.

• Minus one dBTP: the sensible default for anything going to streaming.
• Minus one point five to minus two: dense, bright or heavily limited material, and anything that will be transcoded more than once.
• Minus two or lower: broadcast and film deliverables, where the specification usually states it explicitly and rejection is automated.
• Zero: nothing. There is no delivery context that benefits from it.

The one real exception

Vinyl and other analogue-destined masters do not have this constraint in the same form, because there is no digital reconstruction stage at playback. They have their own, stricter constraints instead — sibilance, low-frequency stereo content, and total duration against groove spacing — and a cutting engineer will tell you what they want.

For everything else, the rule is short. Set the ceiling below zero, measure it with a true peak meter after the last thing you do to the file, and stop worrying about it. It is one number, it takes ten seconds to check, and it is the difference between the record arriving as you made it and arriving with grit on every loud consonant.