Exploring deep learning architectures for urban sound classification

Abstract

Urban sound monitoring is an essential part of city planning as it helps to understand the acoustic environment and its impact on citizens. Machine learning is a valuable tool in this field, providing important insights into the city's soundscape and supporting informed decision making, which can be useful for implementing noise mitigation measures and ensuring regulation compliance using sound source identification. Low-cost acoustic sensors have proven to be a highly effective way to acquire input data, offering scalable and widespread acoustical coverage of a city. The objective of this project is to improve upon existing sound classification models by comparing the performance of deep learning architectures such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and the hybrid Recurrent Convolutional Neural Network (RCNN) approach. This study will also explore more sophisticated architectures, which incorporate dual channel feature extraction and temporal-frequency attention mechanisms (TFCNN). By evaluating the effectiveness of these models, the project aims to drive the field forward and support the development of smart, sustainable cities. The goal is to provide decision-makers with the insights and tools they need to create a better acoustic environment for city dwellers.