Sound Classification and Processing of Urban Environments: A Systematic Literature Review.

Area of Science:

• Environmental acoustics
• Machine learning
• Smart city technologies

Background:

• Audio recognition is vital for smart city applications like security and autonomous vehicles.
• Urban sound environments present complex, unstructured audio data, posing significant classification challenges.
• Existing research focuses on improving sound event detection amidst noise and irrelevant sounds.

Approach:

• This literature review synthesizes recent advancements in audio recognition for urban environments.
• Key factors analyzed include Deep Learning (DL) architectures, attention mechanisms, data augmentation, and pretraining.
• The review identifies state-of-the-art methodologies and their performance on benchmark datasets.

Key Points:

• Deep Learning architectures, particularly DenseNet-161 with ImageNet pretraining, show high performance.
• Data augmentation techniques (NA-1, NA-2) are crucial for enhancing model robustness.
• Top results achieved over 99% accuracy on datasets like ESC-10.

Conclusions:

• Advanced DL models and augmentation strategies are effective for urban sound classification.
• The practical deployment and validated effectiveness of these models in real-world smart city scenarios remain largely unaddressed.
• Further research is needed to bridge the gap between laboratory performance and real-world applicability.