Thesis Abstract
Gestural Improvisation is a type of free improvisation that aims to create short and long-term relationships in the acoustic parameters of music, such as pitch and intensity. The gestural improvisation style is closely related to the concept of real-time composition. This work describes a set of experiments based on signal processing methods and machine learning techniques that are applied to gestural improvisation music. These methods are commonly applied to other styles of music and employed in music information retrieval and computational musicology, among other elds. As a case-study, a set of twenty-one multi-channel recordings of music performed by the ve-member gestural improvisation group EnsAmble Crumble is analyzed.

The theoretical framework presented includes a state-of-the-art scenario of computer improvisatory systems as part of an overview of musical concepts related to improvisation and structure. A section explaining the tools and technologies employed for analysis is given as a reference and a guide. The analyzed data is described in detail and a brief description of the author's previous experience in the development of gestural improvisation and improvisatory systems is also given. Finally, the results and conclusions of the analysis are presented with examples from the study: the behavior of one instrument during the collective performance, the global interaction of one instrument across all of the recordings, and the interaction of all the members across all of the improvisations.

In conclusion, basic rules of interaction were found that can be employed in the generation of computational models of musical improvisation. Such models may be integrated into virtual musical partners with a better musical intelligence than the systems currently available. A set of appendixes includes three articles that complement the work. The appendix contains the description of the Moz-Art-Global-Art project.

Download thesis PDF