| Résumé |
We propose to apply the principle of interactive real-time corpus-based concatenative synthesis to search in effects or instrument sound databases, which becomes content-based navigation in a space of descriptors and categories. This surpasses existing approaches of presenting the sound database first in a hierarchy given by metadata, and then letting the user listen to the remaining list of responses. It is based on three scalable algorithms and novel concepts for efficient visualisation and interaction: Fast similarity-based search by a kD-Tree in the high-dimensional descriptor space, a mass--spring model for layout, efficient dimensionality reduction for visualisation by hybrid multi-dimensional scaling, and novel modes for interaction in a 2D representation of the descriptor space such as filtering, tiling, and fluent navigation by zoom and pan, supported by an efficient 3-tier visualisation architecture. The algorithms are implemented and tested as C-libraries and Max/MSP externals within a prototype sound exploration application. |
