Low-Dimensional Feature Representation for Instrument Identification

Mizuki IHARA, Shin-ichi MAEDA, Kazushi IKEDA, Shin ISHII
2012 SICE Journal of Control Measurement and System Integration  
For monophonic music instrument identification, various feature extraction and selection methods have been proposed. One of the issues toward instrument identification is that the same spectrum is not always observed even in the same instrument due to the difference of the recording condition. Therefore, it is important to find non-redundant instrument-specific features that maintain information essential for high-quality instrument identification to apply them to various instrumental music
more » ... yses. For such a dimensionality reduction method, the authors propose the utilization of linear projection methods: local Fisher discriminant analysis (LFDA) and LFDA combined with principal component analysis (PCA). After experimentally clarifying that raw power spectra are actually good for instrument classification, the authors reduced the feature dimensionality by LFDA or by PCA followed by LFDA (PCA-LFDA). The reduced features achieved reasonably high identification performance that was comparable or higher than those by the power spectra and those achieved by other existing studies. These results demonstrated that our LFDA and PCA-LFDA can successfully extract low-dimensional instrument features that maintain the characteristic information of the instruments.
doi:10.9746/jcmsi.5.249 fatcat:4p2fahuq5rcirhjtbvc4qiohgy