Robust Generative Audio Quality Assessment: Disentangling Quality from Spurious Correlations

TL;DR

This paper introduces a domain adversarial training approach to improve the robustness of audio quality assessment models by disentangling true perceptual quality from dataset-specific biases, enhancing generalization to unseen data.

Contribution

It systematically explores domain definition strategies for disentangling quality from nuisance factors, demonstrating their effectiveness in improving correlation with human ratings.

Findings

Domain adversarial training reduces acoustic biases.

Aspect-specific domain strategies outperform static priors.

Models generalize better to unseen generative scenarios.

Abstract

The rapid proliferation of AI-Generated Content (AIGC) has necessitated robust metrics for perceptual quality assessment. However, automatic Mean Opinion Score (MOS) prediction models are often compromised by data scarcity, predisposing them to learn spurious correlations-- such as dataset-specific acoustic signatures-- rather than generalized quality features. To address this, we leverage domain adversarial training (DAT) to disentangle true quality perception from these nuisance factors. Unlike prior works that rely on static domain priors, we systematically investigate domain definition strategies ranging from explicit metadata-driven labels to implicit data-driven clusters. Our findings reveal that there is no "one-size-fits-all" domain definition; instead, the optimal strategy is highly dependent on the specific MOS aspect being evaluated. Experimental results demonstrate that our aspect-specific domain strategy effectively mitigates acoustic biases, significantly improving correlation with human ratings and achieving superior generalization on unseen generative scenarios.