- author = "Maximos A. Kaliakatsos-Papakostas",
- title = "Computational intelligence methods for automated musical analysis and synthesis",
- school = "University of Patras",
- year = "2014",
- address = "Greece",
- keywords = "genetic algorithms, genetic programming, Music features, Computational intelligence, Automated music synthesis, Automated music analysis, Intelligent music composition",
-
URL = "
http://hdl.handle.net/10442/hedi/37332",
-
DOI = "
doi:10.12681/eadd/37332",
- size = "xxvii + 293 pages",
- abstract = "The PhD thesis at hand discusses the employment of computational intelligence in music, attempting to humbly commit a minimal contribution to the deep history of studies that relate music to mathematics. The three cornerstones upon which the thesis at hand is founded, discuss the employment of computational intelligence methods for a) the examination of musical-mathematical features towards classifying, identifying and characterising music content, b) intelligent music composition based on musical-mathematical features and c) interactive intelligent music composition and further developments. While at a first glance these three parts seem unrelated, their common keystone is the music-mathematical features and the role that these features play towards developing computational intelligence models which at some extent simulate the human perception of music. The fact that all the research channels that are presented in this thesis, are finally led to a single stream, becomes evident in the final chapter of the thesis (Chapter 9) where the music-mathematical features, the intelligent music composition and the interactive music composition are embodied in an innovative system that is thoroughly described. Additionally, a main concern of the studies that comprise this thesis was the presentation of objective, detailed and unbiased results, achieved through exhaustive experimental processes, many of which were by themselves innovative. The latter comment intents to highlight the different approach that the research in this thesis follows, in comparison to the most common approaches concerning the presentation of experimental results for automatic music composition methods - which simply include small score or audio parts of automatically composed music.The first part of the thesis includes the Chapters 2 and 3, where the categorisation of music pieces in symbolic form is examined, as well as the identification and characterisation of music recordings. Aim of this part is on the hand to present the rich quality of information that can be extracted by several pitch class space-related features regarding human perception of music, while on the other hand to pinpoint the effectiveness of computational intelligence methods as tools to extract the aforementioned rich information. The first parts contribution is primarily the presentation of novel methodologies that achieve effective categorisation of music pieces per composer or style, identify the content of drums recordings and characterise the content of recorded pieces by recognising locations of composition key changes. An additionally contribution of this part is the presentation and study of the principal chroma eigenspace.The second part encompasses Chapters 4, 5, 6 and 7, which discuss the contribution of this thesis in intelligent music composition. Specifically, the contribution of Chapter 4 includes a proposed categorisation of intelligent music composition methods based on their intended result, proposing their segregation to unsupervised, supervised and interactive intelligent music composition methodologies. Through this categorisation, an introduction to the subsequent chapters is achieved, which mainly discuss supervised intelligent music composition based on music-mathematical features for the generation of rhythmic sequences (Chapter 5), tonal sequences (Chapter 6), as well as integrated synthesis through the concept of horizontal orchestration replication and intelligent improviser accompaniment (Chapter 7). The results of the presented studies in this part constitute of exhausted research processes that examine different compositional aspects of the proposed methodologies, revealing their strengths and weaknesses over other methodologies presented in the literature.In the third part the interactive systems that were studied in the thesis are presented, not only by analysing the algorithmic development of the underlying methodologies, but mostly focussing on matters that pertain to the human perception and intelligent music composition. Specifically, in the beginning of Chapter 8 an innovative system is presented that evolves mathematical functions interactively (through user ratings), through genetic programming. Aim of this system is the generation and evolution of waveforms that sound more pleasant to the user, according to hers/his subjective criteria. This system allowed the proposition to obtain information about several audio features of the melodies in different evolutionary stages - from non evolved and low rated melodies to evolved and highly rated ones - in order to study whether these features incorporate indications about the aesthetic integrity of a melody. This system was also used towards the development of fitness-adaptive genetic operators, which, combined with the risk factor parameter, gave the user additional control over the evolutionary process, alleviating user fatigue at a considerable extent. The third part, along with the research conducted in the context of this thesis, concludes with Chapter 9, where an interactive evolutionary intelligent music composition system is presented, that combines almost all research presented in the thesis up to that point. This chapter includes also several innovative research propositions in many levels: the core concept, the implementation and the experimental process. The core concept discusses the evolution of music-mathematical features that describe a melody, rather than evolving the melody per se (or the model that generates it). The implementation incorporated two levels of serial evolution: the upper level of feature evolution and the lower level of supervised intelligent music composition, with novel algorithms in both levels. Finally, the experimental process that was developed, in the context of which automatic raters that simulate human behaviour was proposed, allowed a completely subjective evaluation of the systems capabilities, regarding its convergence to the optimal melodies of the users subjective preference.",
-
notes = "in Greek
ND = 37332
Supervisor: Michael N. Vrahatis",
