Dilated-Gated Convolutional Neural Network with A New Loss Function on Sound Event Detection

• Published in: 2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) pp. 1491 - 1495
• Main Authors: He, Ke-Xin, Zhang, Wei-Qiang, Liu, Jia, Liu, Yao
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.11.2019
• Subjects: Convolution; Convolutional neural networks; Kernel; Logic gates; Recurrent neural networks; Task analysis; Training
• ISSN: 2640-0103
• DOI: 10.1109/APSIPAASC47483.2019.9023308

In this paper, we propose a new method for rare sound event detection. Compared with conventional Convolutional Recurrent Neural Network (CRNN), we devise a Dilated-Gated Convolutional Neural Network (DGCNN) to improve the detection accuracy as well as computational efficiency. Furthermore, we propose a new loss function. Since frame-level predictions will be post processed to get final prediction, continuous false alarm frames will lead to more insertion errors than single false alarm frame. So we adopt a discriminative penalty term to the loss function to reduce insertion errors. Our method is tested on the dataset of Detection and Classification of Acoustic Scenes and Events (DCASE) 2017 Challenge task 2. Our model can achieve an F-score of 91.3% and error rate of 0.16 on the evaluation dataset while baseline achieves an F-score of 87.5% and error rate of 0.23.