Guest lecture: Tree structured representation of music for music information retrieval
on Tuesday 14th August at 14:15 Exactum, room C222
David Rizo Valero, University of Alicante, Department of Software and Computing Systems
Abstract: "Monophonic melodic similarity has been tackled so far basically using three approaches: string distances, probabilistic models, and geometric representations. String representation have been the most used to compare melodies. None of the many possible combinations of pitch and duration use to take effective advantage of rhythm information. We propose an alternative tree representation of melodies that inherently incorporates rhythm, as well as standard pitch representations. By applying tree edit distances some good results have been obtained. However those distances have high temporal complexities. I will show some preprocessing of the trees that overcome this problem getting a reasonable trade-off between precision and time. For the experimentation phase of those algorithms there were a lack of big enough corpora. The construction of a convenient corpus for our task led to new problems: the big amount of MIDI files available were multi-track polyphonic songs, and the incorrect tagging of key information at most MIDI files. Two algorithms that try to resolve those problems will be presented: the first guesses which of the tracks in a MIDI file is the melody one, the second extracts the key from a song by using the same kind of tree representation already presented. The tree representations are good at removing ornament notes, by deciding which of the notes in the tree are the most important. Rules to prune the trees presented so far in the are developed on a experimentation basis instead of having a musicological background. Having as target to make those pruning rules more robust, an automatic tonal analysis system has been developed getting promising results. To finish the talk, I will present the extension to polyphonic music of the trees in order to compare polyphonic songs. The results will be compared to those using the geometric methods that are being developed here at the University of Helsinki"
Welcome!