XAI-based Comparison of Input Representations for Audio Event Classification

Different ways of representing audio data lead to fundamentally different decision-making strategies: The study found that models using raw waveforms versus time-frequency spectrograms (visual representations of sound) employ distinct classification approaches, demonstrating that input representation choice is not merely a technical detail but shapes how AI systems "think" about audio.

Explainability tools can reveal whether AI models use human-aligned reasoning: By using Layer-wise Relevance Propagation heatmaps, researchers could verify that models' classification strategies align with what humans consider relevant features for audio event detection, providing a method to validate that AI systems are learning for the "right reasons."

Performance metrics alone are insufficient for evaluating audio AI systems: The paper demonstrates that measuring only classification accuracy misses crucial information about how models make decisions. Understanding the underlying strategies through explainability is necessary to assess robustness and ensure models will generalize reliably to new situations.

Explainable AI enables more informed design choices: Rather than selecting input representations based solely on which produces the highest accuracy scores, the XAI approach allows developers to choose representations that produce more robust and interpretable decision-making processes, supporting more responsible AI deployment.