SPolDB: an audio dataset for artificial intelligence-based identification of noise pollutants

Worldwide technological advancements have introduced machines for various tasks, from simple spice mixing to heavy drilling. Many of these devices produce multifarious sounds which are often over tolerable limits, catering to noise pollution. This rising threat affects biodiversity and human health, with visible impacts like the disappearance of birds and health complications. Artificial Intelligence (AI) has been employed to detect sound pollutants, aiding in soundscape mapping for better urban planning and biodiversity protection. However, data scarcity remains a challenge for training machine learning models. We present SPolDB, a comprehensive sound pollution dataset with 54 classes from indoor and outdoor sources, featuring over 133,000 clips of varying lengths. The dataset was composed using recordings in the natural ambiance as well as sourcing audio clips of real-world scenarios. Baseline results are reported using an established handcrafted feature-based approach (Mel Frequency Cepstral Coefficient + Random Forest/Multi-layered Perceptron) and deep learning approach. The highest balanced accuracy (to mitigate bias introduced by class imbalance) of 82.9% was obtained on the test set (20% data for each class) leveraging a customized lightweight Convolutional Neural Network architecture named XZ-Net along with Mel Spectrogram.