A Walkthrough to Airborne Acoustic Source’s Hidden Patterns in Lower and Higher Dimensional Space

Recent AI advances, like Transformers, Conformers, Residual Neural Networks, and EfficientNets, have exceeded expectations. These models’ effectiveness relies significantly on input data. Traditional signal processing and machine learning (ML) techniques, which use key features as inputs, remain promising. However, the shift to using raw signals in end-to-end systems can hinder performance, particularly when diverse classes are grouped together. This study highlights the advantages of using a broad set of temporal, spectral, and spectro-temporal features to capture the characteristics of airborne acoustic signals. The insights from this research can help optimize model performance by fine-tuning input features according to specific requirements.