Automatic genre classification of Latin American music using characteristic rhythmic patterns

In the majority of musical genres, music is basically composed of repetitive rhythmic structures (patterns). Especially in Latin American music, particular styles can be distinguished through characteristics of these patterns. Therefore, the aim of the present work is the automatic classification of musical genres from Latin America using automatically extracted rhythmic patterns. The approach is based on setting up a knowledge base that consists of typical reference patterns for each genre. To obtain a tempo independent pattern representation, we apply both the scale transform and the log-lag autocorrelation function. Different distance measures were evaluated to measure the similarity between unknown patterns and reference patterns. Various tests with different preprocessing techniques were performed. For 9 distinct genres, a classification accuracy of 86.7% for tests with synthetic data and 47.9% with real-world music was achieved. In addition, conclusions to rhythmic similarity of particular music styles were drawn. Dealing with non-western music, the work presents an operational method for genre classification in the research field of Computational Ethnomusicology.