Dangerous Sound Detection Using Convolutional Feature Extraction and Temporal Modeling with BiLSTM

Abstract

Dangerous sound detection is essential to improve public safety through automated surveillance systems capable of identifying and classifying multiple hazardous acoustic events. This study proposes a hybrid deep learning framework that integrates a one-dimensional Convolutional Neural Network (1D-CNN) for spatial feature extraction with Bidirectional Long Short-Term Memory (BiLSTM) for temporal sequence modeling. The system is designed for multi-class classification, targeting eight distinct categories of dangerous sounds, including gunshots, explosions, screaming, crying, glass breaking, fire, emergency alarms, and weapon handling. A comprehensive set of audio features, such as mel-spectrograms, MFCCs, chroma, spectral contrast, and temporal descriptors, is extracted to capture diverse spectral, tonal, and temporal characteristics of each class. The model achieves high accuracy while maintaining low training and validation losses, demonstrating strong generalization across classes with varying acoustic similarity. Experimental results confirm the system’s robustness in distinguishing between acoustically similar sounds and its ability to handle class imbalance effectively. The architecture, supported by a structured preprocessing pipeline, is optimized for scalability and real-time deployment in complex urban environments. These findings highlight the potential of combining convolutional and recurrent deep learning techniques for robust, multi-class acoustic event detection, with future work focusing on lightweight model adaptation, expanded datasets, and integration of multimodal contextual information to further enhance performance and operational reliability.