PERCUSSION CLASSIFICATION IN POLYPHONIC AUDIO RECORDINGS USING LOCALIZED SOUND MODELS

Abstract

This paper deals with automatic percussion classification in polyphonic audio recordings, focusing on kick, snare and cymbal sounds. We present a feature-based sound modeling approach that combines general, prior knowledge about the sound characteristics of percussion instrument families (general models) with on-the-fly acquired knowledge of recording-specific sounds (localized models). This way, high classification accuracy can be obtained with remarkably simple sound models. The accuracy is on average around 20% higher than with general models alone.