Environmental Sound Classification Based on CNN Latent Subspaces

In recent years, Convolutional Neural Networks (CNN) have become the standard method for addressing various audio classification problems. However, as CNNs rely on the existence of large amounts of training data, their performance can be limited in scenarios where such data is scarce. In this paper, we propose signal latent subspace (SLS), an alternative sound classification method, achieving competitive results without requiring a huge amount of data. Using two different pre-trained CNN models, we extract two distinct feature sets of a given sound. Then, we model an independent signal latent subspace representation of each feature set, before unifying them on a product Grassmann manifold (PGM) to perform classification. We evaluate our method on the task of Environmental Sound Classification (ESC) through quantitative experiments on UrbanSound and ESC-10 datasets, achieving the accuracy of 85.01% and 88.60% respectively.