Introduction

The problem of sound source identification is not only an academic curiosity on how the human brains work and how to make a computer system which can do the same. A desire for automatic classification of audio materials according to instruments makes the problem a practical one. A number of techniques and features have been experimented with in order to identify the musical instrument from an isolated tone [1] [2]. There have been, however, far fewer works which consider identification problems from polyphonic signals. Among them, Eggink and Brown [3] used energy bands as features, omitting from use in classification the bands which tend to have overlapping spectra. They obtained on average 49% identification rate using five instruments over one pitch range. Time-domain template matching and features related to note onset and spectral distribution were investigated in [4] and [5] respectively for two-tonal mixtures from a set of three different instruments.

In this paper, ISA is used to decompose a mixture into its ``statistically independent'' components and spectral bases, hopefully spanning the subspace of each original source in the mixture. The system does not rely on pitch estimation in contrary to previous systems. Physiologically intuitive features can also be derived from the learned bases for classification which are usually lost when more than one sources are active simultaneously.