A Convolutional Model with Multi-Scale Feature Fusion, Duo Pooling, and Fusion Loss for Acoustic Scene Classification

Acoustic Scene Classification (ASC) aims to identify an environment specific to its unique audio cues. Convolutional Neural Networks (CNN) have become the de-facto baseline in modern ASC systems. In general, the efficiencies of the signal representation and the learning model are crucial to the performance of a convolutional model for ASC. In this paper, these factors are addressed as follows. First, a signal representation method based on multi -scale feature fusion that leverages the representation diversities is introduced. Second, a novel technique to aggregate more learned features in the pooling layer: “duo pooling” - is proposed. Third, a “Fusion loss” that encourages the model to learn enhanced representations is presented. Experimental results on a standard ASC dataset indicate that our model outperforms a baseline with a log - mel spectrogram representation by 82.85 %. In addition, it has a good measure of separability between the individual classes, as indicated by an AUC (Area Under the ROC Curve) score of 0.995.