Quality Adaptive Angular Margin Learning for Respiratory Sound Classification

The article presents QLung, a quality-adaptive angular-margin learning framework designed for respiratory sound classification, which enhances feature generalization by ensuring intra-class compactness and inter-class separability. It introduces a no-reference audio quality margin based on spectral entropy and root-mean-square energy, utilizing a log-scaled angular margin to stabilize training amidst class imbalance. The framework achieves a 2.46% improvement in in-distribution performance on the ICBHI dataset and outperforms existing methods on the SPRSound dataset, making it significant for practitioners focused on robust classification in varying audio quality scenarios.