Using different combination strategies to combine features for audio scene classification based on Convolutional Neural Networks

Abstract Audio scene classification (ASC) which can make good use of audio information for classification is currently a hot topic. In this paper, taking Convolutional Neural Networks (CNN) as the classification model, we have designed three combination strategies to combine different kinds of features for classification. In the first combination strategy, the multi-channel CNN is used to combine different features; in the second combination strategy, different kinds of features are concatenated, and the concatenated features are used as the input of the CNN; in the third combination strategy, new features are first extracted from each kind of feature through CNN, and then these new features are concatenated for classification. Experiments are done on the DCASE2017 datasets to test the effectiveness of the combination strategies, and the experimental results show that the second combination strategy performs best, also the experimental results show that on the whole, combining more features would get better classification results.