Lists
Lists
Meraviglia, Manfredo (2023) Computer generated music: a new empiricial approach. Doctoral thesis, London Metropolitan University.
Genetic algorithms have been widely used for systems of automatic music generation; however, the field lacks research on the aesthetic value, on harmony or musical form and on control over the kind of music generated by such systems.
This thesis outlines investigations into scoring methods based on Zipf's Law and the Golden Ratio in order to successfully score the balance between binary elements (e.g., repetition-variation). These investigations were conducted through the development and implementation of a music-generation system named lntelMusica. The focus was kept on fitness rules that can be applicable to traditional tonal music and experimental music alike. In order to reach aesthetically pleasing results, the generation process mimics a natural composition method where a melodic idea (created by the genetic algorithm) is developed into a full musical piece through variations, complete with harmonisation and arrangement. Each stage of musical development is scored by the Fitness Function in order to select the best candidate to reach the next stage: two-bar ideas are randomly generated, harmonised and scored. The best candidates are then combined into four-bar phrases, these phrases are scored. Phrases are subsequently combined in eight-bar periods, scored and then they are combined into thirty-two bars pieces (or some other length suitable to the selected form). The finished music pieces are scored and the best is selected as output. The fact that genes utilised for the evolution of melodies are generated randomly and not from patterns or variations of pre-existing music allows the system to explore musical spaces unusual in western popular music.
The Fitness Function was applied to three categories of human-composed music: Beatles songs, Alternative Hits (Billboard) and Hits (Billboard) and found to be able to correctly classify these in order of perceived pleasantness.
Finally, the respondents to an evaluation survey found the music composed by the system to be more pleasant than music produced by other comparable systems and more pleasant than the included human-composed samples. Furthermore, respondents holding a degree-level education (or higher), found samples composed by the system with an altered scale pleasant.
The question this thesis poses is not "can a computer be creative?", but rather "how can a computer create and identify something of value?" This thesis then proposes the following answer: a possible method for a computer to identify a valuable melody is measuring its balance of melodic skips to steps against the Golden Ratio and its scale-degree distribution against Zipf's Law.
 |
View Item |