Sparse Representation Frameworks for Acoustic Scene Classification

This work addresses the task of acoustic scene classification (ASC) by using sparse representation frameworks, motivated by the inherent sparseness of audio data. We explore three different sparse representation classification (SRC) frameworks, generating sparse acoustic scene representations. The first two frameworks focus on producing linear and non-linear features respectively. On the other hand, the third framework presents a novel approach-a two-branch deep sparse auto-encoder (DSAE) representation framework that generates non-linear and discriminative features. In the proposed framework, the first branch induces sparsity, while the second focuses on enforcing discrimination within the learned sparse acoustic scene representations. These representations are later used to classify the acoustic scene data into different acoustic scene classes. We also compare the performance of the three sparse frameworks by evaluating them on three ASC datasets. Our results indicate that acoustic scene representations based on DSAE outperform the sparse representations obtained from the other two frameworks. This results in an average performance gain of approximately 8% across all the ASC datasets.