Speaker
Description
Timbre's temporal micro-variability contributes substantially to source identity and sound quality. However, existing temporal representations are either computationally driven or perceptually motivated but reduced to scalar descriptors. Building on a momentary approach to timbre developed in prior work, we present a framework that isolates temporal variability and represents it as a continuous trajectory in a low-dimensional space.Perceptual grounding is realized through asymmetric sampling in time: parallel feature streams are extracted at short (~30 ms) and long (~200 ms) integration windows. We compare wavelet scattering features against log-mel spectrum and MFCC baselines. The streams are embedded with a diffusion map into a geometric representation motivated by the non-metric, relational character of perceptual timbre spaces. Over this embedding we apply two complementary topological summaries: Mapper graphs to expose the qualitative shape and connectivity of variability trajectories, and density-aware persistent homology (DTM filtration) for quantitative signatures comparable across sources. We are not aware of prior work combining perceptual constraints with a geometric representation of within-note timbre variability.We present an initial case study on sustained notes from instruments spanning wind, brass, and string families. Early embeddings show family-consistent groupings, though we treat these observations as illustrative at the time of submission; a complete evaluation is expected in time for the conference. The framework is intended as a foundation for computable, perceptually valuable descriptions of timbre variability.