'Copy and scale' method for doing time-localized MIR estimation: Application to beat-tracking

In this work we propose a 'copy and scale' method based on a Nearest Neighbour paradigm to estimate timelocalized parameters and apply it to the problem of beattracking. The Nearest Neighbour algorithm consists in assigning the information of the closest item of a preannotated database to an unknown target. It can be viewed as a 'copy and paste' method. The 'copy and scale' method we propose consists in 'scaling' this information to adapt it to the properties of the unknown target. In order to represent time-location, we represent the content of an audio signal using a sampled and tempo-normalized complex discrete Fourier transform (DFT) of its onset energy function. This representation is used as the code over which the Nearest Neighbour search is performed. Along each code of the Nearest Neighbour space, we store the corresponding annotated beat-marker positions in a normalized form. A search is then performed for a set of tempo assumptions. Once the closest code and best tempo assumption are found, the normalized beat-markers of the closest item are scaled to this tempo in order to provide the estimation of the beatmarkers of the unknown item. We perform a preliminary evaluation of this method and show that, with such a simple method, we can achieve results comparable to the ones obtained with sophisticated approaches.