MSA-TCN: Robust Urban Suspicious Sound Detection Using Multi-Scale Temporal Convolutions and Dual Attention

Efficient suspicious sound event detection is essential for urban surveillance, smart-city monitoring, and public safety systems; however, many existing deep learning approaches suffer from high computational complexity, unstable optimization, and limited suitability for real-time deployment. This paper proposes a Multi-Scale Attention-based Temporal Convolutional Network (MSA-TCN) for robust urban sound classification. The model operates on MFCC representations and integrates dilated multi-scale temporal convolutions to capture both short impulsive events and long-duration acoustic patterns. To enhance training stability and generalization, spectral normalization is applied to convolutional layers, while channel and spatial attention mechanisms are incorporated to adaptively emphasize discriminative temporal–spectral features. The proposed MSA-TCN is evaluated on the UrbanSound8K dataset using stratified 80/10/10 train–validation–test splits and compared against Dense and LSTM baseline models under identical experimental conditions. MSA-TCN achieves a test accuracy of 92.34% and a macro-averaged F1-score of 0.92, outperforming the Dense and LSTM baselines by 3.48% and 2.09%, respectively, while using only 128K parameters. Comprehensive ablation studies demonstrate that spectral normalization and multi-scale temporal modeling are the most critical components, each contributing approximately 2–2.5 percentage points to test accuracy and significantly improving convergence stability. Attention mechanisms primarily enhance training efficiency and class-specific discrimination rather than absolute accuracy. Confusion-matrix and per-class analyses confirm strong recall for safety-critical events such as gunshots and sirens, while training dynamics reveal stable convergence and efficient learning. These results validate MSA-TCN as a lightweight, robust, and effective framework for real-time urban sound classification in public safety applications.