This dissertation is focused on the development of generative audio systems - a term used describe generative music systems that generate both formal structure and synthesized audio content from the same audio-rate computational process. In other words, a system wherein the synthesis and organizational processes are inseparable and operate at the sample level.
First, a series of generative software systems are described. These systems each employ a different method to create generativity and, though they are not strictly generative audio systems, they lay important groundwork for the rest of the discussion as ideas from and contributions to the fields of generative algorithmic composition, computational aesthetics, music information dynamics, and digital signal processing are introduced.
Second, the dissertation investigates the use of a novel signal processing technique in which time-varying allpass filters are placed into feedback networks, producing synthesis structures capable of yielding interesting emergent sonic behaviors. Ideas from the field of computational aesthetics are employed to allow a large system built from these synthesis structures to become “aesthetically aware.” Many theories about computational aesthetics center around a favorable balance between order and complexity in a stimulus - a successful artistic work is neither too orderly nor too complex. Using a model of human perception based on the “mere exposure” effect, which describes how listener appreciation and boredom change as they experience repeated exposure to a stimulus, the AAS-4 system autonomously determines when and how to modify its own parameters to avoid repetitions that may lead to boredom in listeners.
The dissertation concludes with objective analysis of the generative system by considering the complexity of its output from an information-theoretic perspective. It was found that the generative audio system described here is capable of producing output with equivalent complexity to that of real-world musical examples. It is also shown that the level of complexity in the generated audio and real-world examples falls in-between the low complexity of silence and sinusoids and the maximal complexity of white noise, corresponding with the theories from computational aesthetics. Future directions of this work are also described. Two appendices describing related topics that would disrupt the flow of the dissertation are included.
