Back to list

Automatic Extraction of Drum Sounds from Music Recordings

Authors Hüsler, G.
Year 2017
Thesis Type Audio Engineering project
Topic Audio Signal Processing
Abstract The aim of this work is to implement an algorithm that automatically locates,extracts and classi es isolated drum sounds from music recordings. The use of drum sounds taken from external music recordings is a wide spread technique in modern popular music, also called sampling. For musicians locating and slicing these relatively short sound snippets can be very time consuming, especially when working with big music collections. Therefore, such a system could help musicians by making it easier to create drum libraries from their personal music collections. The procedure to establish the considered algorithm can roughly be divided into four stages. In the fi rst stage potential sections in the waveform are selected using crest factor and root mean square and in the second stage beat tracking is used to segment the sections into individual sounds. Subsequently the sounds are classi ed into the categories non percussive/harmonic,kick drum, snare and hi-hat. For the classi cation step a set of temporal and spectral features are calculated and reduced in dimensionality by using linear disciminant analysis. For the actual classi cation maximum likelihood method is used. Labeled training data was gathered from several drum libraries and breakbeat compilations. An overall classi cation rate of 89% was achieved, which may be further improves by using more training data.
Supervisors Sontacchi, A.