Comparing Onset Detection & Perceptual Attack Time

Accurate performance timing is associated with the perceptual attack time (PAT) of notes, rather than their physical or perceptual onsets (PhOT, POT). Since manual annotation of PAT for analysis is both time-consuming and impractical for real-time applications, automatic transcription is desirable. However, computational methods for onset detection in audio signals are conventionally measured against PhOT or POT data. This paper describes a comparison between PAT and onset detection data to assess whether in some circumstances they are similar enough to be equivalent, or whether additional models for PAT-PhOT difference are always necessary. Eight published onset algorithms, and one commercial system, were tested with five onset types in short monophonic sequences. Ground truth was established by multiple human transcription of the audio for PATs using rhythm adjustment with synchronous presentation, and parameters for each detection algorithm manually adjusted to produce the maximum agreement with the ground truth. Results indicate that for percussive attacks, a number of algorithms produce data close to or within the limits of human agreement and therefore may be substituted for PATs, while for non-percussive sounds corrective measures are necessary to match detector outputs to human estimates.