Article
A Role-Authenticated Interpretable AI System for Multi-Tiered Urban Noise Monitoring in Smart Cities
Urban environments generate complex acoustic landscapes with overlapping sound events such as sirens, horns, and traffic noise, making automated monitoring for smart cities, public safety, and noise pollution control highly challenging. Traditional sound classification systems rely on hand-crafted features like Mel-Frequency Cepstral Coefficients (MFCCs) and log-Mel spectrograms, combined with models such as Support Vector Machines (SVMs), Random Forests (RF), or shallow Convolutional Neural Networks (CNNs), achieving only modest accuracy. Emerging from early 2010s research and benchmarked on datasets like UrbanSound8K, these approaches suffer key limitations: inability to capture semantic audio understanding in overlapping conditions, neglect of multi-label correlations, lack of interpretability in deep models, and dependence on command-line interfaces that limit usability for non-experts. This research proposes a Whisper-powered multi-task urban sound classification system with an integrated Graphical User Interface (GUI) built using Tkinter and secured with rolebased authentication (Lightning Memory-Mapped Database (LMDB) and SHA-256 hashing). It leverages OpenAI Whisper-base as a feature extractor, generating robust representations through mean pooling of encoder hidden states from class-organized audio data. These features are used to train four interpretable models: Boosted Rules Classifier (BRC), Hierarchical Structural (HS) Tree Classifier, Sparse Linear Integer Model (SLIM), and Marginal Shrinkage Linear Trees (MSLT), enabling dualtask classification of primary categories and subcategories. The system enhances transparency, usability, and performance by combining transformer-based representations with interpretable machine learning. It democratizes access to explainable artificial intelligence (AI) for urban monitoring, enabling secure, visual, and multi-task sound analysis for real-world smart city applications.
Full Text Attachment
