5 February 2015
Statistical Modelling of Music Using Multiple Viewpoint Systems
Meeting Room 10, 2nd Floor, JLB
12:30pm - 13:45pm
Mr Raymond Whorley
A statistical model of music is constructed by machine learning from a corpus, such that (ideally) novel music in the style of the corpus can be generated from the model. Viewpoints are able to model attributes of the musical surface, such as pitch and note duration; and also more abstract attributes, such as interval and scale degree. They can be linked together to model conjunctions of attributes, and they can be threaded to model long-range dependencies. We shall use a tiny melodic corpus to examine how a viewpoint model is created from n-gram models of different order, and how viewpoint models are themselves combined (and so on, to the top of the model hierarchy). An important part of the machine learning process, automatic viewpoint selection (guided by the information theoretic measure cross-entropy), will also be examined.
Once the fundamentals have been covered, we shall move on to issues relating to the domains (alphabets) of viewpoints, while at the same time introducing and comparing several versions of the multiple viewpoint framework for harmony. In particular, we shall examine the importance of a robust procedure to construct viewpoint domains; the impact on computational complexity of large domains within the harmonic framework; and how that impact can be mitigated. Finally, we shall investigate, with examples, the relationship between cross-entropy and harmonic correctness (according to some of the more general rules of harmony); and how a simple modification to random sampling enables the rapid generation of harmonisations containing few rule violations.
Ray Whorley has been involved with music making from an early age, and has an ‘A’ level in music. He learned the piano and ’cello; has sung with several choirs; and has written orchestrated settings of translations of Arabic and Persian texts.
He initially attended the University of Leeds, gaining a BSc (Hons) in chemical engineering in 1980. He developed a career in this field, working for a variety of companies including BOC Cryoplants, Kvaerner Process (UK), BNFL Engineering, Snamprogetti and MW Kellogg. Having been interested in computing from his school days, he took all opportunities to create computer tools; for example, for project person-hour estimation and for the selection and sizing of filters and separators.
In 2002, he initiated a switch to computing by doing the Diploma in Computer Science at the University of Cambridge. He took a rule-based approach in his project and dissertation “Automatic Harmonisation of a Given Melody.” Shortly afterwards, he gained some commercial programming experience at Cambridge Systems Associates, a company specialising in financial software. Later still, he was given a departmental studentship to do a PhD in computer science at Goldsmiths, University of London, where he was supervised by Christophe Rhodes, Geraint Wiggins and Marcus Pearce. The title of his thesis was “The Construction and Evaluation of Statistical Models of Melody and Harmony,” with an emphasis on harmony. Since the award of his PhD in late 2013, he has been collaborating with Darrell Conklin (University of the Basque Country and IKERBASQUE) on improving techniques for sampling from statistical models.
Save to your Calendar