Clustering Spectral Filters for Extensible Feature Extraction in Musical Instrument Classification
Patrick Donnelly, John W. Sheppard · 2014
We propose a technique of training models for feature extraction using prior expectation of regions of impor-tance in an instrument’s timbre. Over a dataset of train-ing examples, we extract significant spectral peaks, cal-culate their ratio to fundamental frequency, and use k-means clustering to identify a set of windows of spec-tral prominence for each instrument. These windows are used to extract amplitude values from training data to use as features in classification tasks. We test this ap-proach on two databases of 17 instruments, cross evalu-ate between datasets, and compare with MFCC features.