Environment sound classification using an attention-based residual neural network

Complexity of environmental sounds impose numerous challenges for their classification. The performance of Environmental Sound Classification (ESC) depends greatly on how good the feature extraction technique employed to extract generic and prototypical features from a sound is. The presence of silent and semantically irrelevant frames is ubiquitous during the classification of environmental sounds. To deal with such issues that persist in environmental sound classification, we introduce a novel attention-based deep model that supports focusing on semantically relevant frames. The proposed attention guided deep model efficiently learns spatio-temporal relationships that exist in the spectrogram of a signal. The efficacy of the proposed method is evaluated on two widely used Environmental Sound Classification datasets: ESC-10 and DCASE 2019 Task-1(A) datasets. The experiments performed and their results demonstrate that the proposed method yields comparable performance to state-of-the-art techniques. We obtained improvements of 11.50% and 19.50% in accuracy as compared to the accuracy of the baseline models of the ESC-10 and DCASE 2019 Task-1(A) datasets respectively. To support the attention outcomes that have focused on relevant regions, visual analysis of the attention feature map has also been presented. The resultant attention feature map conveys that the model focuses only on the spectrogram's semantically relevant regions while skipping the irrelevant regions. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

The study introduces a novel attention-based deep model for environmental sound classification, which efficiently learns the spatio-temporal relationships in the spectrogram and achieves comparable performance to state-of-the-art techniques.