Bergeron, M. (2010). Structured polyphonic patterns. (Unpublished Doctoral thesis, City University London)
- Accepted Version
Download (1MB) | Preview
The present dissertation develops, applies and evaluates a novel method for the representation and retrieval of patterns in musical data. The method supports the typical polyphonic patterns that one finds in music theory textbooks. Most current computational methods to musical patterns are restricted to monophony (one melody at a time). The Structured Polyphonic Patterns method (SPP) applies to the general case of polyphonic music, where many melodies may unfold concurrently. Pattern components are conjunctions of features which encode properties of musical events, or relations that they form with other events. Relations between events that overlap in time but are not simultaneous are supported, enabling patterns to express key temporal relations of polyphonic music. Patterns are formed by joining and layering pattern components into sequences (horizontal structures) and layers (vertical structures). A layer specifies voicing in an abstract way, and the exploration of different voice permutations is handled automatically. The SPP method also provides a mechanism for defining new features. We evaluate SPP by developing a small catalog of musicologically relevant queries and analyzing the results on four corpora: 185 chorale harmonizations by J.S. Bach, Mozart Symphony no. 40, a small set of piano pieces by Chopin, and a collection of folk songs containing more than 8000 pieces – in addition to its size, demonstrating the scalability of the method, that latter corpus is interesting as it shows that SPP is also usable for monophony. Examining several corpora allows us to establish that some polyphonic patterns constitute salient properties of a corpus: they are over-represented in one corpus by comparison to the others. In addition, the queries we develop demonstrate that the SPP method possesses sufficient expressiveness to capture important music-theoretic notions. At the same time, we show how the method is more restrictive than some existing polyphonic pattern representations, hence providing a better approximation of the expressive power required for polyphonic patterns. It is a better candidate representation for music data mining, a difficult problem that has received significant attention for the monophonic case, but limited attention for the more general polyphonic case.
|Item Type:||Thesis (Doctoral)|
|Divisions:||School of Informatics > Department of Computing|
Actions (login required)
Downloads per month over past year