Interpreting deep urban sound classification using Layer-wise Relevance Propagation

Explainability is critical for safety-sensitive applications: The researchers emphasize that for a system designed to assist hearing-impaired drivers, "clear etiology justifying and interpreting model predictions comprise a strong requirement," demonstrating that AI transparency isn't optional but essential when systems affect user safety.

Multiple perspectives improve trustworthiness: By analyzing the same audio using two different feature representations (Mel and constant-Q spectrograms), the team identified which frequency content showed "extremely discriminative information" across both methods, providing stronger evidence for why the model makes specific decisions rather than relying on a single interpretation.

Layer-wise Relevance Propagation makes neural networks interpretable: The paper demonstrates a practical technique for opening the "black box" of deep neural networks by tracing which parts of the input (specific sound frequencies) most influenced each prediction, making the model's reasoning transparent to users and developers.

Explainability enables accountability: By documenting which audio features drive classifications, this framework allows developers and users to verify the model is using appropriate reasoning (e.g., actual traffic sounds) rather than spurious correlations, supporting responsible deployment in real-world applications.