Abstract: Acoustic scene classification (ASC) aims to distinguish between different acoustic environments and is a technology which can be used by smart devices for contextualization and personalization. Standard algorithms exploit hand-crafted features which are unlikely to offer the best potential for reliable classification. This paper reports the first application of convolutional neural networks (CNNs) to ASC, an approach which learns discriminant features automatically from spectral representations of raw acoustic data. A principal influence on performance comes from the specific convolutional filters which can be adjusted to capture different spectrotemporal, recurrent acoustic structure. The proposed CNN approach is shown to outperform a Gaussian mixture model baseline for the DCASE 2016 database even though training data is sparse.

Download: