Audio peak
Also known as: transient, level peak
The loudest instant in a stretch of sound, useful both as a level warning and as a marker for where something physically happened.
Updated
What it means
An audio peak is a momentary maximum in the waveform. Technically it matters because a peak too close to the ceiling distorts, which is why level meters show peaks rather than averages.
Editorially it matters for a different reason. On action footage a peak is usually an event: a board landing, a door closing, a shout, an impact against water.
That makes peaks a cheap source of truth about timing. The picture may be ambiguous about when something happened, while the sound rarely is, which is why Audio event detection leans on them.
How it works in practice
Peak detection is arithmetic rather than anything learned, so it runs over a whole upload for almost nothing and produces candidate moments before any model is involved.
Peaks are strongest as corroboration. A motion spike and an audio peak at the same instant is a real event, while either alone is a guess, which is how a Hallucination gets caught.
Levels want headroom rather than maximum loudness. Leaving a few decibels under the ceiling means a later stage can raise the whole mix without anything clipping, which is the difference between a track that survives mastering and one that does not.
What to watch out for
Wind produces peaks that mean nothing. A gust against an exposed microphone reads as loud and sudden without anything having happened, which is why Wind noise has to be filtered out before peaks are trusted, and why a Windscreen (dead cat) changes the signal as much as any processing does.
A peak also says nothing about what made it. A dropped camera and a landed trick look identical in the waveform, so the signal needs picture to interpret it.
Chasing peaks during capture leads to a quiet, thin recording. Setting levels so nothing ever peaks means most of the track sits far too low to use.
How RawClip handles it
Loud instants in the camera audio are used as candidate moments and as corroboration for what the picture suggests, so an impact that shows in both sound and movement is treated as real while one that shows in neither is dropped.