多任务实时声音事件检测卷积模型与复合数据扩增.

Most of the existing sound event detection researches analyzes from recorded audio, and these models had massive parameters with low efficiency, which is not suitable for real-time detection. This paper proposed an efficient convolutional neural networks model for multi-task and real-time sound event detection. The model considered triggering and detecting task as multi-task learning mission, and its structure combined densely connection, ghost module and SE block. Moreover, this paper proposed a data augmentation method that combined with audio shifting, random cropping and SpecAugment. The experimental results show that the mean prediction accuracies of the model on ESC-10 and Urbansound8K datasets are at least 2% higher than latest baseline models, but also has better robustness as well as less parameters and memory. Additionally, multi-task learning saves computation, and due to the reusing of feature maps, the model can quickly and accurately feedback the results. Compared with the traditional data augmentation methods, the proposed model achieves better performance and robustness. [ABSTRACT FROM AUTHOR]