Hybrid and Depth Separable Convolutional Neural Network Approaches for an Efficient and Accurate Siren Recognition With Resource Limited Devices

Emergency vehicles, such as ambulances, fire trucks, and police cars, play a crucial role in road traffic management. While these vehicles typically use sirens for identification, many drivers rely primarily on visual cues. This reliance presents challenges in scenarios where the emergency vehicles are obscured or outside the observer’s field of vision, posing significant difficulties for individuals with hearing impairments or reduced ability to perceive ambient sound. This paper explores the feasibility of efficiently detecting emergency vehicle sirens on resource limited devices using innovative neural network models. Specifically, we propose (i) a hybrid model by integrating a one‐dimensional convolution neural network (CNN) with a gated recurrent unit (GRU) and (ii) a depth separable convolutional neural network to demonstrate that they are efficient models suitable for deployment on resource constrained devices such as microcontrollers (MCUs) with a Cortex‐M core, FPGAs, and TPUs. Experimental results indicate that the hybrid model achieves 98% average accuracy with 434.5 μJ energy consumption and 12.32 ms latency on STM32F411RE, enabling 52.8 h battery life in continuous monitoring applications. The depth‐wise separable convolutional neural network (DSCNN) variants provide energy‐efficient alternatives with 85% accuracy consuming 28% less energy, enabling deployment‐specific optimization for safety‐critical versus autonomous monitoring scenarios. Detailed ablation studies and Grad‐CAM visualizations establish the interpretability and novelty of the proposed architectures for transparent emergency detection on MCUs.