MIREX 2005: Symbolic Genre classification with an ensemble of parametric and lazy classifiers

The symbolic genre classification algorithm submited to the MIREX (Music Information Retrieval Exchange) 2005 is described here. Our algorithm uses a combination of k-nearest neighbors and Bayesian classifiers trained with different sets of statistical descriptors extracted from melody tracks extracted from MIDI files. It is aimed at classifying melodies by genre. The statistical descriptors describe pitch, note duration, silence duration, and rhythmic properties of the melody. The set of descriptors is invariant to transposition or tempo scaling and deliberately contains no information based on metadata, such as instrumentation or text data. Descriptors consist mainly in counters, range, average, standard deviation of musical properties. Each track is reduced to a monophonic sequence of notes, prior to the extraction of descriptors. Classifiers are trained independently using different subsets of descriptors. The resulting models are combined using a majority vote scheme. In order to select a melody track from a MIDI file, a model of ’melody track’ previously trained is applied to each track in a MIDI file. The most probable melody track is selected and used as an instance for the different classifier ensembles. Therefore each MIDI file is classified using information based on a single track.