CLASSIFICATION INTO MUSICAL GENRES USING A RHYTHMIC KERNEL

Publication Type:

Conference Paper

Source:

SMC Conference 2004 (2004)

URL:

files/proceedings/2004/P21.pdf

Abstract:

Beginning with the question on how to determine the genre of a music piece, we elaborate on the representation of rhythm for the classification into genres. The aim of such classification differs in principle from that of traditional Music Informa tion Retrieval algorithms. First, we formalise the rhythmic representation of music fragments. This formalism is then used to construct a similarity function called kernel. To allow the discrete comparison of rhythmic fragments, a pre-processing step in the algorithm computes a common quantization unit among the input data. A simple injective mapping into N allows the kernel to employ the Euclidean dot product. A small database of jazz, classical and rock fragments is used in an implementation of a Support Vector Machine. The issues that arise with different time signatures are analysed. Finally, we share some early results of the experiments comparing the three genres, showing that rhythm conveys good information for classification, within the conditions of the experiment.