Comparison of Hand-Crafted and Deep Features Towards Explainable AI at the Edge for Analysis of Audio Scenes

This paper explores explainable Artificial Intelligence (XAI) for pre-trained audio models and audio source counting. An XAI pipeline is proposed that utilizes pretrained audio models as feature extractors to investigate the efficacy of hand-crafted acoustic features in replicating the learned representations of the deep features. The approach involves two phases: evaluating similarity between hand-crafted and deep features using the linear centered kernel alignment algorithm, followed by classification analysis using Multi-Layer Perceptrons and select hand-crafted features. The results show this approach has implications for edge computing and XAI, where we demonstrate a resource-constrained neural network classifier, under 10 KB in size, that uses interpretable hand-crafted features, exhibits less than a 4% decrease in accuracy compared to a benchmark model <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$136\times$</tex> larger.