SunEcho: An Optimized Deep Learning Model for Real-Time Urban Environmental Sound Classification

Abstract Noise pollution represents a growing public health crisis, with 466 million people worldwide experiencing disabling hearing loss in 2019, projected to reach 700 million by 2050. Approximately 80\% of affected individuals reside in low- and middle-income countries, where limited capacity to identify and monitor noise sources exacerbates the problem. This paper presents an optimized urban noise classification system designed for deployment on resource-constrained edge devices to enable continuous environmental monitoring. We investigated four convolutional neural network architectures—SunEcho, Spec-CNN, AlexNet, and LeNet-5—using the Sunbird/urban-noise-uganda-61k dataset with two input representations: log-mel spectrograms and YAMNet embeddings. Models were evaluated under fine-grained (19-class) and categorical (6-class) taxonomies. Spectrogram-based inputs consistently outperformed embeddings across all architectures, with the custom SunEcho model achieving optimal performance: 89\% categorical accuracy and 81\% fine-grained validation accuracy. The system provides city authorities in developing regions with an accessible, deployable tool for evidence-based noise source identification and mitigation strategies to improve public health outcomes.