Music Retrieval

Contact: Remco Veltkamp


Music is digitised in several different forms, the most important of which are audio information and encoded notation. In principle, both of these can be used for retrieval purposes, but even in the easier case—retrieval from encoded notation—it is hard to attain meaningful output. One reason is that music notation maps musical events onto a two-dimensional space, only one of which (pitch) is generally employed for retrieval. The proportions of false positives and false negatives are both generally unacceptably high in pith-only searches. The obvious solution would be to take other parameters into account as well (at least duration), but methods for this are still in their infancy. An additional problem is that users generally want to search for similarity, not identity. A measure for musical similarity still needs to be found that satisfies computational and cognitive requirements.


The objective of the Orpheus project is to approach the following three fundamental issues in music retrieval:
  • Matching. Given a query theme or melody, determine if it is contained in collection of (possibly longer) melodies or complete works.
  • Polyphonic matching. Given a query melody, determine is it is contained in a polyphonic piece of music.
  • Indexing. Given a large collection of music, build a data structure to efficiently search for those pieces similar to the query.

C-Minor (Cognition-based Music Information Notation Oriented Retrieval)
The objective of the C-Minor project is to approach the following three fundamental issues in music retrieval:
  • Feature extraction. Given a piece of music, or a musical fragment such as a theme or melody, compute a number of features that are relevant to music perception and cognition.
  • Matching. Design a similarity measure between those features, that models well perceptual similarity, and design an efficient algorithms to compute it.
  • Relevance feedback. Given a selection from the collection, with positive and negative assessments from the user, find music that is more similar to the positively labeled music, and more dissimilar to the negatively labeled music.


The following people are involved in this research area:

webmaster: Remco Veltkamp