A Proximity Grid Optimization Method to Improve Audio Search for Sound Design

Sound designers organize their sound libraries either with dedicated applications (often featuring spreadsheet views), or with default file browsers. Content-based research applications have been favoring cloud-like similarity layouts. We propose a solution combining the advantages of these: after feature extraction and dimension reduction (Studentt Stochastic Neighbor Embedding), we apply a proximity grid, optimized to preserve nearest neighborhoods between the adjacent cells. By counting direct vertical / horizontal / diagonal neighbors, we compare this solution over a standard layout: a grid ordered by filename. Our evaluation is performed on subsets of the One Laptop Per Child sound library, either selected by thematic folders, or filtered by tag. We also compare 3 layouts (grid by filename without visual icons, with visual icons, and proximity grid) by a user evaluation through known-item search tasks. This optimization method can serve as a human-readable metric for the comparison of dimension reduction techniques.