Slicing by Transients

Transients, for our purposes, are short, almost click-like portions of a sound. They are distinct from onsets because they don't necessarily correspond with the start of some event like a note or percussive strike.

Cleary there is a close relation though, and the starts of notes, drum hits and other events very often do have audible transients associated with them. However, if we think about very textural sounds–such as pouring rice onto a hard surface, or scrunching up tin foil–we might hear these as singular events that are composed of lots of transients.

Modelling transients is quite involved, and research in this area has been driven, in particular, by work in audio restoration (de-clicking, for example) and audio coding (for making compression algorithms like MP3 or AAC). One common tactic among the diverse approaches in the literature is to treat transients as brief moments of unpredictability in a signal, and use the errors from some other way of modeling the signal as indicators for transients.

Using this approach to slice signals is comparatively computationally intensive, but can yield very precise results. Because the algorithm works in the time domain, the locations of estimated transients can be very precise (to within a few samples of where a human might locate it in a DAW), and the process can run with a very small latency.

Also, because looking for transients is not the same as looking for the start of events, like notes, this kind of slicing technique can work for very fine-grained segments, especially in more textural sounds.