Use Case Scenario: Identifying Themes and Variations

Short Description

Given a group of tracks representing a theme and a set of variations, using either the theme or any of the variations as query, find the other tracks in the group.

(Probably only feasible for sets of variations before, roughly, Beethoven, where the length of variations and their harmonic structure is likely to be close to that of the theme.)

Beneficiaries

Musicologists, but also others in that a generalised track-based harmonic model could allow rapid pre-filtering of large collections and restrain more exhaustive search to tracks with overall 'harmonic content similarity' (or 'harmonic palette/language'?).

Task related to composer style-recognition (ADBComposerClassifier); OMRAS 1 [MIDI] experiments showed how such a model does in general rank 'non-relevant' works by the same composer higher than might be expected in such searches.

Also relates to cover song detection (e.g. ADBDifferentPerformerSameWorkClassifier)

Task Description using fftExtract and AudioDB

Might involve the development of a meta-feature derived from chroma that incorporates harmonic similarity. (It is an open question how the 'smoothing' inherent in chroma features derived from real-world audio will interfere with a harmonic model of this sort.)

Also experiment with using audioDB/LSH for matching whole-track Markov-like models (of order > 0, and thus of high dimensionality) to capture harmonic-transition information.

Data Set and data set preparation

Ground Truth

I have a small test set (Bach Goldberg variations; Handel Harmonious Blacksmith variations; artificially degraded Bach chorales) which can be used as a basis for comparison with former OMRAS (MIDI-based) work done with Jeremy Pickens (ISMIR 2002).

Experimental Method

Results Tables and Figures

Conclusions

-- TimCrawford - 17 Sep 2008