Technical Issues

In this project we applied the evolution of text processing in computers to the practice of computer-aided composition. The basic idea was to create an abstraction layer between the compositional and the computational process. Very few synthesis methods have been able to create such a layer which gives the composer the ability to think globally about music. With its efficiency for creating sound, FM synthesis[5] was probably a strong factor in the commercialization of synthesizers; however, it falls short of providing any tool which is musically intuitive. In my opinion, that is why our ears are able to recognize the sound of FM rather easily and get bored by it. I believe that CHANT, which uses the FOF synthesis method developed by Rodet[37], has been the most successful technique for providing a musical synthesis method for the composer (See Barrier[1] and Harvey[17]).

We defined a synthesis method which made no distinction between the micro and macro-structures of music. The synthesis parameters are defined as a hierarchy of structures which can contain recursive elements. The system can create self-similar or self-affine sounds from a number of different point of views (e.g., pitch fluctuation, amplitude fluctuation, the shape of the spectrogram, and the way different layers of sound or music are faded in). A simple language was developed for specification of the hierarchy. The system proved to be able to create extremely complex results with very simple structures; however, not every result was musically interesting to us. Most of the research work with the system was to search for structures which resulted in musically interesting sounds. These structures showed a certain versatility that, through very little change, could create new sounds which were different from the original results but still remained interesting to our ears. Thus, the relationship in the structures defined a certain class of sound in the system which could be tailored for a specific purpose. A piece was composed using the system which shows that it is possible to create sounds with specific intentions.

The concept of self-similarity was used because since self-referentiality was formally introduced to me by ``Gödel, Escher, Bach''[18], I have been rediscovering it in many unexpected contexts. When we combined some very simple computer science ideas such as programmability, hierarchy, and functionality to what we knew of computer music, self-similarity had developed itself in the design of the system by unifying the different perceptual levels in the model. In our search, a sense of duality was discovered in the traditional way that two concepts were treated: one in the treatment of sound and music, and the other in the technical treatment of random and deterministic signals.

noise was studied and a simple analysis of many pieces which we had access to in MIDI format was conducted. A signal was extracted from these pieces and almost all the spectrum of the extracted signals showed a slope close to that of noise, which means that the signal is neither random nor too correlated. noise falls on the border between the signals which we treat as random (for which we use statistical methods to study) and deterministic signals (for which we use very precise functions). This class of signals creates some technical problems by the fact that a signal with a spectrum shows a scale invariant auto-correlation, which in turn means that there are certain correlations among all levels of the signal (i.e. micro or macro-structures).