From Rhythm Patterns to Perceived Tempo

There are manyMIR applications for which we would like to be able to determine the perceived tempo of a song automatically. However, automatic tempo extraction itself is still an open problem. In general there are two tempo extraction methods, either based on the estimation of interonset intervals or based on self similarity computations. To predict a tempo the most significant time-lag or the most significant inter-onset-interval is used. We propose to use existing rhythm patterns and reformulate the tempo extraction problem in terms of a nearest neighbor classification problem. Our experiments, based on three different datasets, show that this novel approach performs at least comparably to state-of-the-art tempo extraction algorithms and could be useful to get a deeper insight into the relation between perceived tempo and rhythm patterns.