A perceptually based onset detector for real-time and offline audio parsing

Use of the perceptually determined Bark frequency scale is investigated in an implementation of spectral difference onset detection. The study is carried out using a new external object for onset detection in Pure Data called bark∼, and contrasted with a similar object called bonk∼. While the filterbanks used in both of these objects feature more detailed resolution in low frequency bands, bark∼’s reliance on the Bark scale is intended to make transitional filter spacing from low to high frequency more appropriately gradual. A weighting curve is also applied in order to account for variabilities associated with equal loudness curves. In three performance evaluations involving piano note onsets, it is shown that bark∼ produced fewer erroneous reports and was able to capture onsets that bonk∼ (similarly configured) failed to report.