Detection of Piano Pedaling Techniques on the Sustain Pedal

Automatic detection of piano pedalling techniques is challenging as it is comprised of subtle nuances of piano timbres. In this paper, we address this problem on single notes using decision-tree-based support vector machines. Features are extracted from harmonics and residuals based on physical acoustics considerations and signal observations. We consider four distinct pedalling techniques on the sustain pedal (anticipatory full, anticipatory half, legato full and legato half pedalling) and create a new isolated-note dataset consisting of different pitches and velocities for each pedalling technique plus notes played without pedal. Our results using cross-validation trails show the effectiveness of the designed features and the trained classifiers for discriminating pedalling techniques.