Onset detection locates the start of new sound events in an audio signal by finding rapid increases in energy or spectral flux

Onset detection identifies the moments at which new musical events begin in an audio stream. Methods fall into time-domain (energy envelope peaks), spectral (flux: sum of positive spectral differences between frames), and phase-based approaches. A rectified spectral flux function highlights frames where new frequency content appears regardless of global loudness. A peak-picking stage then identifies local maxima above an adaptive threshold. Onset detection underpins many higher-level tasks: beat tracking, tempo estimation, segmentation, and rhythm quantisation. Challenges include soft onsets, polyphony (overlapping onsets), and varying note density.

Examples

In Ableton Live, Warp markers use onset detection to identify beats. In SuperCollider: Onsets.kr UGen. Chop in Strudel subdivides a sample at detected transients.

Assessment

Why is spectral flux a more robust onset detector than pure energy-envelope for sounds like bowed strings (which have slow attacks)? What makes onset detection harder in a polyphonic ensemble recording than a solo instrument?