Pattern Recognition -MAS 622-


A method to classify sounds of musical instruments (single monophonic notes) is introduced. The classification is done using in parallel two sets of perceptual features extracted from the sounds. The models used are a mixture of gaussians, whose parameters where found by training over a database of target sound families. The feature extraction procedure, model training and model usage for classification are explained. An implementation and results of the method are shown and discussed.

Download PDF document

Feature Selection

Spectral Contour


Cepstral Coefficients

Gaussian Mixture model

Ricardo A. Garcia MIT Media Lab 2001