Graduate School and Research Center in Digital Sciences

Acoustic scene classification using convolutional neural networks

Battaglino, Daniele; Lepauloux, Ludovic; Evans, Nicholas

DCASE 2016, Workshop on Detection and Classification of Acoustic Scenes and Events, September 3rd, 2016, Budapest, Hungary

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.

Document Bibtex

Title:Acoustic scene classification using convolutional neural networks
Keywords:acoustic scene classification, convolutional neural networks, local binary patterns, spectrogram
Department:Digital Security
Eurecom ref:4982
Bibtex: @inproceedings{EURECOM+4982, year = {2016}, title = {{A}coustic scene classification using convolutional neural networks}, author = {{B}attaglino, {D}aniele and {L}epauloux, {L}udovic and {E}vans, {N}icholas}, booktitle = {{DCASE} 2016, {W}orkshop on {D}etection and {C}lassification of {A}coustic {S}cenes and {E}vents, {S}eptember 3rd, 2016, {B}udapest, {H}ungary}, address = {{B}udapest, {HUNGARY}}, month = {09}, url = {} }
See also: