Automatic Labeling of Unpitched Percussion Sounds
Perfecto Herrera, Amaury Dehamel, Fabien Gouyon · 2003
We present a large-scale study on the automatic classification of sounds from percussion instruments. Different subsets of temporal and spectral descriptors (up to 208) are used as features that several learning systems exploit to learn class partitions. More than thirty different classes of acoustic and synthetic instruments and near twothousand different isolated sounds (i.e. not mixed with other ones) have been tested with ten-fold or holdout crossvalidation. The best performance can be achieved with Kernel Density estimation (15% of errors), although boosted rule systems yielded similar figures . Multidimensional scaling of the classes provides a graphical and conceptual representation of the relationships between sound classes, and facilitates the explanation of some types of errors.