Freesound Drum Sets Using Unconventional Sounds
The aim of this thesis is to investigate whether we can use sounds deriving from objects not originally designed for musical purposes to construct novel drum sets capable of efficiently functioning in a musical context. We tackle the subject by developing a system trained with standard drum kit sounds, built with the purpose of classifying random percussive-like sounds from unconventional objects under drum set instruments' labels. To this end, we collect two datasets, each including sounds of
... ncluding sounds of the two aforementioned types, both subsequently made freely accessible in Freesound. The former dataset is created in compliance with our drum set taxonomy which we hold to be among the most detailed in relevant literature. Our system, implemented in both the Weka workbench and python's scikit-learn Machine Learning library receives as an input selected sound features extracted with the use of one of Essentia's out-of-box extractors. After comparing a multitude of different algorithms, we concluded upon the use of k-Nearest Neighbor as the most suitable for the task at hand. After extensive experimentation we establish a list of nine descriptors to be used in our classifier, while achieving accuracy rates (calculated with tenfold cross-validation) close to the highest presented in relevant state-of-the-art studies (reaching over 90%). The efficiency of our model is tested through an experiment involving human participants, where a series of sequences performed by unconventional drum sets is rated on the basis of their ability to substitute conventional drum sets, their similarity with the later as well as the listeners' appreciation thereof. Finally, we produce as proof of concept a simple step sequencer producing drum patterns with sounds from our "unconventional" Freesound dataset, with the aim to include it in Freesound Labs.