Novel dataset and model for restroom sound event classification.

This study presents a novel privacy-preserving deep learning framework for accurately classifying fine-grained hygiene and water-usage events in restroom environments. Leveraging a comprehensive, curated dataset comprising approximately 460 min of stereo audio recordings from five acoustically diverse bathrooms, our method robustly identifies 11 distinct events, including nuanced variations in faucet counts and flow rates, toilet flushing, and handwashing activities. Stereo audio inputs were transformed into triple-channel Mel spectrograms using an adaptive one-dimensional convolutional neural network (1D-CNN), dynamically synthesizing spatial cues to enhance discriminative power. Extensive experimentation identified the RegNetY-008 architecture as the most effective backbone, further improved by employing a semi-supervised learning strategy via pseudo-labeling and targeted data augmentation techniques such as XY masking and horizontal CutMix. The proposed ensemble model, combining RegNetY-008 networks with complementary third-channel generation strategies, achieved outstanding generalization performance, yielding an accuracy of 97.8% and macro-averaged F1-score of 0.966 across acoustically distinct test environments. Our publicly available dataset addresses critical gaps in existing resources, promoting future research in intelligent, privacy-conscious restroom monitoring.