PctN: A perceptually grounded indicator for quantifying natural sound exposure for ambient noise mitigation

Quantifying natural sound exposure is critical for controlling ambient noise within urban contexts, yet isolating natural sound components within complex urban acoustic environments remains a significant technical challenge, despite the human auditory system’s inherent ability to perform discrimination. To bridge this gap, this study proposes a perceptually grounded indicator PctN (Percentage of Natural Sound Sources) to identify natural sound exposure, and develops a machine-learning framework for its automated quantification. Developed from extensive field recordings and controlled auditory experiments, PctN quantifies natural sound exposure on a normalized scale from 0 to 1. Correlation analyses reveal significant associations between PctN and a range of acoustic, psychoacoustic, and Mel spectrogram features, confirming Mel spectrograms as the most informative determinant of PctN. Three machine-learning architectures were evaluated for PctN prediction, among which a Convolutional Recurrent Neural Network (CRNN) trained on Mel spectrograms achieved the highest prediction accuracy (0.83). The framework was further enhanced using a Transformer architecture to improve performance in real-world deployments. Implemented as an automated measurement instrument, the approach was validated through precise identification of the dawn chorus, demonstrating its capability to capture dynamic variations in natural sound exposure within complex urban sound environments. This research establishes a perceptual-informatics sensing framework for quantifying acoustic naturalness for managing ambient noise.