Convolutional-Recurrent Neural Network with the Tensor Fusion Mechanism for Acoustic Scene Classification

Acoustic scene classification (ASC) is one of the key fields of artificial intelligence. Due to the short duration of scene audio features, the existing deep learning network cannot fully capture information in short-term audio. In this regard, a convolutional-recurrent neural network with the tensor fusion mechanism (CRN-FM) is proposed for ASC. Each audio is divided into fixed-length segments, and the spectral features are extracted from the segment audio as the input. Then, a convolutional neural network (CNN) is used to obtain time-frequency related information, and long short-term memory (LSTM) is used to obtain time-related details. When receiving the output of the high-level features by the two modules, the designed tensor fusion attention layer fuses different tensors according to the difference in information saturation. Finally, a SoftMax classifier is used to classify scenes. Experimental results on DCASE 2018 and 2019 ASC datasets demonstrate the effectiveness of the proposed approach.