Deep Segment-Attentive Network for Altered-Engine Recognition.

Altered engine recognition is of significant research and application prospects since it is a sub-task of environmental sound classification and helps detect altered vehicles for law enforcement. In this work, we propose a simple but effective framework based on a sound embedding network with an attention mechanism to identify noise emitted from vehicles with altered engines. Real-world vehicle sound data were collected at a number of streets with different amounts of traffic in Hong Kong. In particular, we developed a proprietary dataset consisting of the environmental and engine sounds through manual segmentation and annotation, and we found that the attention mechanism can help emphasize the informative segments in the frame-level embeddings outputted by the embedding network. We also demonstrate the effectiveness of the proposed attention mechanism on the ESC-50 and UrbanSound8K datasets. The experimental results show that our method achieves remarkable performance on UrbanSound8K and outperforms the baseline on ESC-50 by more than 10%.