Drum Detection from Polyphonic Audio via Detailed Analysis of the Time Frequency Domain

This publication presents a method for the automatic detection and classification of three distinct drum instruments in real world musical signals. The regarded instruments are kick, snare and hi-hat as agreed by the participants of the contest category Audio Drum Detection within the 2nd Annual Music Information Retrieval Evaluation eXchange (MIREX 2005). There are two challenging issues inherent to drum sound recognition in polyphonic music. The first problem is that the drum sound itself can vary greatly within the same instrument class, due to playing techniques, recording situation and electronic effects. The second apparent problem is the interference and masking with all other instruments sounding simultaneously with the drum in a musical signal, making it difficult to reliably detect occurrences of a certain drum type. The method outlined here achieves a solution to these problems by extending a source separation approach described in earlier publications with spectrogram templates and a more elaborate classification approach. Performance results of the system are given by the outcomes of the Audio Drum Detection contest within the MIREX 2005.