18 Octubre 2011
Joan Serrà

In this talk I will present an overview of the work done within my PhD, in which we studied several computational approaches for the identification of versions of the same musical piece (near-duplicate music recordings). In particular, I'll present our two main approaches for the automatic retrieval of music versions, which solely use audio-content information. The first approach is based on nonlinear time series analysis techniques and it is currently the most accurate system of its class according to MIREX, an international evaluation framework for music processing algorithms. The second approach is based on prediction concepts and it alleviates the computational costs of the first one (at the expense of being less accurate). I'll also introduce a straightforward but effective post-processing step that enhances the accuracy of music retrieval systems based on the so-called query-by-example paradigm.