Optimized YamNet Model for Early Detection of Emergency Situations in Smart Safety Systems

—Auditory perception is crucial for recognizing environmental cues, especially in emergencies. However, noise and overlapping signals can hinder the identification of hazardous sounds, compromising safety responses. Therefore, developing an advanced sound recognition system capable of accurately identifying and analyzing these signals to mitigate such limitations is essential. To address this, we leveraged YAMNet, a deep learning model based on the MobileNetV1 architecture, which employs depthwise separable convolutions to achieve high computational efficiency. This approach allowed for real-time sound recognition in complex acoustic environments while maintaining low latency, making it ideal for deployment in smart safety systems. Building on YAMNet’s feature extraction capabilities, we integrated a custom model with Dense layers, fine-tuning it to better classify important emergency sounds such as alarms, gunshots, and distress signals. Experimental results showed that the system achieved 91.64% accuracy and an F1-score of 0.89, effectively identifying hazardous signals in noisy settings. These findings demonstrated the system’s potential to enhance emergency response efficiency in real-time warning applications.