Benchmarking Time-localized Explanations for Audio Classification Models

TL;DR

The paper addresses the challenge of evaluating time-localized explanations for audio classification models where ground-truth explanations are scarce. It introduces a perturbation-based benchmarking framework that masks audio in 100 ms segments, records model responses via the log-odds $\log(P_c/(1-P_c))$, and trains surrogate models to assign per-segment importances. The evaluation uses datasets with time-aligned event annotations (e.g., KWS, AudioSet subsets, synthetic Drums) and reports averaged AUC; findings show that random-forest surrogates often yield stronger explanations and that some tasks (like KWS) can reach near-perfect AUC. It also demonstrates practical utility for uncovering spurious correlations (Clever-Hans) and provides public datasets and code for ongoing benchmarking.

Abstract

Most modern approaches for audio processing are opaque, in the sense that they do not provide an explanation for their decisions. For this reason, various methods have been proposed to explain the outputs generated by these models. Good explanations can result in interesting insights about the data or the model, as well as increase trust in the system. Unfortunately, evaluating the quality of explanations is far from trivial since, for most tasks, there is no clear ground truth explanation to use as reference. In this work, we propose a benchmark for time-localized explanations for audio classification models that uses time annotations of target events as a proxy for ground truth explanations. We use this benchmark to systematically optimize and compare various approaches for model-agnostic post-hoc explanation, obtaining, in some cases, close to perfect explanations. Finally, we illustrate the utility of the explanations for uncovering spurious correlations.

Figures (4)

• Figure 1: Procedure to produce explanations for a given black box model and input signal. First, a set of masks are created. Then, each mask is used to create a perturbed version of the input signal, which is fed into the model. The resulting scores, together with the corresponding masks, are used to estimate the impact of masking each segment of the input signal.
• Figure 2: Average AUC for the datasets in our benchmark for different explanation approaches. The whiskers correspond to the 95% confidence interval obtained by bootstrapping the test samples ferrer2024goodpracticesevaluationmachine.
• Figure 3: Average AUC on the Drums dataset as only the first part of the ground truth for each event is preserved.
• Figure 4: Example explanation for an IEMOCAP "happy" sample with a summed cough (in red) for two models: one trained with the original IEMOCAP samples and one trained with the corrupted version that contains coughs for the "happy" class.