| Abstract | Automatic classification of musical instrument is an old research topic in music
information retrieval. In this work we address the problem of the classification using
musical instrument single note samples from Freesound and we put the accent
on the content analysis of the sound and how those content information are connected
to the physical characteristics of each instrument. We build a taxonomy
based on instrument families and mode of excitation. The musical instruments
play a central role in this work and the studies on timbre are used as a methodological
base to apply feature selection to our complete set of descriptors, the aim
of this is to find which descriptors are relevant to describe a specific instrument.
The machine learning then is used as an instrument to evaluate our choices, to
identify weakness and problem in the current implementation of audio descriptors.
|
