Learning from Soft-Computing Methods on Abnormalities in Audio Data

In our research we deal with polyphonic audio data, containing layered sounds of representing various timbres. Real audio recordings, musical instrument sounds of definite pitch, and artificial sounds of definite and indefinite pitch were applied in this research. Our experiments included preparing training and testing data, as well as classification of these data. In this paper we describe how results obtained from classification allowed us to discover abnormalities in the data, then adjust the data accordingly, and improve the classification results.