8–12 Sept 2026
Europe/Vienna timezone

Topological Features for Environmental Sound Classification

FA2026/122
11 Sept 2026, 11:00
20m
Saal 3 (Messe Congress Graz)

Saal 3

Messe Congress Graz

A09 Machine learning and artificial intelligence in acoustics A09.00 Machine learning and artificial intelligence in acoustics

Speaker

Jiwon Seo (Fraunhofer IDMT)

Description

In this work, we investigate whether topological features derived from persistent homology of Takens embeddings can complement conventional acoustic features for environmental sound classification on the ESC-50 dataset. A compact statistical summary of persistence diagrams is extracted from short overlapping segments, aggregated at the recording level, and combined with conventional acoustic features. Experiments show that fusing topological and acoustic descriptors consistently improves classification performance across multiple classifiers. With a Random Forest classifier, the proposed fusion achieves 48.5% accuracy, an absolute improvement of 7.6 percentage points over the baseline. With a multilayer perceptron, the improvement reaches 12.1 percentage points, yielding 55.1% accuracy. A dimensionality-matched control experiment confirms that these gains reflect informational content, not extended dimensions. This work provides empirical evidence that persistent homology captures structural properties of environmental sounds that are complementary to conventional acoustic features.

Authors

Jiwon Seo (Fraunhofer IDMT) Sascha Grollmisch (Fraunhofer IDMT) Jaehyun Lee (LOAS Inc.) Sukjin Hong (LOAS Inc.) Jooyoung Hahn (FNSPE, CTU in Prague)

Presentation materials