Instrument Identification in Monophonic Music Using Spectral Information
Mizuki Ihara, Shin‐ichi Maeda, Shin Ishii · 2007
Various kinds of feature sets have been proposed to represent characteristics of musical instruments. While those feature sets have been chosen in a rather heuristic way, in this study, we demonstrate that the log-power spectrum suffices to represent characteristics that are essential to identifying instruments. For efficient encoding of instrument characteristics, we then reduce the number of features by applying the well-known dimension reduction techniques: principal component analysis (PCA) and linear discriminant analysis (LDA). For the classification of eight instruments, the features obtained by applying PCA-LDA to the log-power spectrum performed very well in comparison to existing methods with a recognition rate of 91% with as few as ten features.