Cross-associating unlabelled timbre distributions to create expressive musical mappings
Proceedings of the First Workshop on Applications of Pattern Analysis, PMLR 11:28-35, 2010.
In timbre remapping applications such as concatenative synthesis, an audio signal is used as a template, and a mapping process derives control data for some audio synthesis algorithm such that it produces a new audio signal approximating the perceived trajectory of the original sound. Timbre is a multidimensional attribute with interactions between dimensions, and the control and synthesised signals typically represent sounds with different timbral ranges, so it is non-trivial to design a search process which makes best use of the timbral variety available in the synthesiser. We first discuss our preliminary work applying standard machine-learning techniques for this purpose (PCA, self-organising maps), and the reasons they were not satisfactory. We then describe a novel regression-tree technique which learns associations between unlabelled multidimensional timbre distributions.