Musical note and instrument classification with likelihood-frequency-time analysis and support vector machines

In this paper, we analyze the classification performance of a likelihood-frequency-time (LiFT) analysis designed for partial tracking and automatic transcription of music using support vector machines. The LiFT analysis is based on constant-Q filtering of signals with a filter-bank designed to filter 24 quarter-tone frequencies of an octave. Using the LiFT information, features are extracted from the isolated note samples and classification of instruments and notes is performed with linear, polynomial and radial basis function kernels. Correct classification ratios are obtained for 19 instrument and 36 notes.