Robust Generative Audio Quality Assessment: Disentangling Quality from Spurious Correlations

2026-03-17

AI Summary

This work addresses the vulnerability of existing automatic audio quality assessment models to data scarcity, which often leads them to learn spurious acoustic correlations tied to specific datasets rather than genuine perceptual quality characteristics. To mitigate this, the authors propose a disentanglement framework based on Domain-Adversarial Training (DAT), systematically exploring diverse domain definition strategies—from explicit metadata to implicit clustering—to separate quality-relevant features from confounding factors. A key insight is that no universally optimal domain partitioning exists; instead, the choice of strategy should be adaptively tailored to different Mean Opinion Score (MOS) dimensions. Experimental results demonstrate that the proposed approach significantly improves correlation with human ratings and exhibits superior generalization to unseen audio generation scenarios.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Application Domains: Humanities & Computational Social Science
• Knowledge Representation and Reasoning: Qualitative Reasoning

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Web Mining and Content Analysis: Web data quality in the era of algorithmically-generated content

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.

Problem

• audio quality assessment
• spurious correlations
• AI-Generated Content
• Mean Opinion Score
• generalization

Innovation

• domain adversarial training
• disentangled representation
• audio quality assessment
• spurious correlations
• aspect-specific domain strategy