Signal latent subspace: A new representation for environmental sound classification

Abstract

In this study, we propose Signal Latent Subspace (SLS), a flexible method that classifies environmental sound events using the subspace representations of latent features obtained from various neural network-based models. Our main goal is to leverage the high expressiveness of neural networks while retaining the advantages of subspace representation, such as its robustness to noise and ability to work under small sample size (SSS) conditions. We also propose an ensemble strategy native to the subspace representation, to achieve increased performance and reduce the generalization error. We do this through product Grassmann manifold (PGM), resulting in SLS-PGM. Each subspace constructed from latent features of a network can be seen as a point on a factor Grassmann manifold (GM) of a neural network; through PGM, it is possible to unify factor manifolds into a singular representation, and perform classification through a similarity metric on the manifold. We further improve SLS and SLS-PGM in two ways: (1) by using generalized difference subspace (GDS) projection to address the lack of between-class discrimination of subspace representation and (2) by leveraging finetuning regimes to better adapt neural network models to the ESC task. We evaluate our proposed methods, factoring various neural networks, on ESC-10, ESC-50 and UrbanSound environmental sound datasets, and provide extensive ablation experiments and notes for practical use.