Application of missing feature theory to the recognition of musical instruments in polyphonic audio

A system for musical instrument recognition based on a Gaussian Mixture Model (GMM) classifier is introduced. To enable instrument recognition when more than one sound is present at the same time, ideas from missing feature theory are incorporated. Specifically, frequency regions that are dominated by energy from an interfering tone are marked as unreliable and excluded from the classification process. The approach has been evaluated on clean and noisy monophonic recordings, and on combinations of two instrument sounds. These included random chords made from two isolated notes and combinations of two realistic phrases taken from commercially available compact discs. Classification results were generally good, not only when the decision between reliable and unreliable features was based on the knowledge of the clean signal, but also when it was solely based on the harmonic overtone series of the interfering sound.