Robust Deep Feature Extraction Method for Acoustic Scene Classification

Abstract

In recent years, increasing number of acoustic scene classification (ASC) methods are based on deep learning models. In these models, the extraction of robust deep feature plays an important role on the classification accuracy. However the complex combination of acoustic phenomena in an acoustic scene results in overlapping of the analysis features, which degrades the performance of ASC. To enhance the compactness of feature and fit the multi-classification task, we explored the data label learning for deep feature extraction. And we combined the method of label smoothing(LS) and the additive margin softmax loss (AM-softmax) to extract deep feature based on VGG-style deep neural network. The comparison experiments show that the best classification results are obtained by the proposed method, which accuracy on ESC-50 dataset is 81.9%, which is beyond human performance.