Compose - Compute. Computer generation and classification of music through operations research methods

Date: 12 December 2014

Venue: University of Antwerp, Promotiezaal Grauwzusters - Lange Sint-Annastraat 7 - 2000 Antwerp

PhD candidate: Dorien Herremans

Principal investigator: Prof. dr. Kenneth Sörensen

Short description: PhD defense Dorien Herremans - Faculty of Applied Economics

Abstract: Can a computer compose music? Can a computer give us insight into what makes a dance song a hit? These questions are examined in this research, by applying quantitative methods from the domain of operations research to problems from music.
In the first part, a music generation system is developed that can generate a continuous stream of classical music. The music adheres to the strict counterpoint rules, a musical style that was developed in the 17th century and that is at the basis of contemporary Western music. The system is based on a variable neighbourhood search algorithm and was implemented as an Android app [FuX]. The name of the app refers to Johan FuX, the 17th century composer that is considered as the founder of counterpoint. The app is freely available in the Google Play store.
In the second part of this research models are created that give more insight into a certain music style. This enables us to generate music automatically based on a certain style, without having a formal description of the style from music theory. Musical style characteristics of three composers where extracted by scanning a large collection of existing musical pieces. Based on this data a model was created that allows us to classify musical pieces per composer (Bach, Beethoven en Haydn). This model was plugged into the FuX app, which enables the generation of music with characteristics of a certain composer. The user of the system can configure the proportion in which the generated music contains the musical characteristics of the three composers. Bach, but with a hint of Haydn; 50% Beethoven and 50% Bach, every combination is possible.
Complex Markov models were also made that describe the counterpoint style and music for bagana, and Ethiopian lyre. Based on these models, different evaluation metrics were developed that can be used by the music generation system to generate music in a certain style. This was combined with an efficient technique to generate cyclic, structured music.
In the last part of this research, data mining was used to build classification models that could predict if a dance song is going to be a top 10 hit versus a lower listed position. This system was implemented as a free online tool (http://antor.uantwerpen.be/dance, which enables the user to upload their own mp3. The system returns the probability that the song will be a top 10 hit.