Acoustic Scene Classification Based on Additive Margin Softmax
Kun Yao, Jibin Yang, Xiongwei Zhang, Changyan Zheng, Xin Zeng · 2019
In the acoustic scene classification task, different categories of scenes may contain the same acoustic events and the sound characteristics of the same category of scenes may also be significantly different, which results in poor classification performance. In order to solve this problem, a deep scene classification model based on convolutional neural network (CNN) is proposed with the additive margin softmax (AM-softmax) function as the measurement loss. The proposed model is able to acquire more categorical representative features of acoustic samples. The test results on the environmental sounds classification datasets ESC-10 and ESC- 50 show that the classification accuracy of the model is improved greatly compared with other methods based on deep learning, which achieve 94.75% and 80.25% respectively.