Training of Classifiers for the Recognition of Musical Instrument Dominating in the Same-Pitch Mix

Abstract

Preparing a database to train classifiers for identification of musical instruments in audio files is very important, especially in a case of sounds of the same pitch, when a dominating instrument is most difficult to identify. Since it is infeasible to prepare a data set representing all possible ever recorded mixes, we had to reduce the number of sounds in our research to a reasonable size. In this paper, our data set represents sounds of selected instruments of the same octave, with additions of artificial sounds of broadband spectra for training, and additions of sounds of other instruments for testing purposes. We tested various levels of added sounds taking into consideration only equal steps in logarithmic scale which are more suitable for amplitude comparison than linear one. Additionally, since musical instruments can be classified hierarchically, experiments for groups of instruments representing particular nodes of such hierarchy have been also performed. The set-up of training and testing sets, as well as experiments on classification of the instrument dominating in the sound file, are presented and discussed in this paper.