Robust Generative Audio Quality Assessment: Disentangling Quality from Spurious Correlations

A research team from Taiwan has published a paper, 'Robust Generative Audio Quality Assessment: Disentangling Quality from Spurious Correlations,' addressing a critical flaw in evaluating AI-generated audio. As tools for creating speech, music, and sound effects proliferate, we need reliable, automated metrics to judge their perceptual quality, typically predicted as a Mean Opinion Score (MOS). Current models, however, are often trained on limited datasets, causing them to learn incidental 'acoustic signatures' of the training data—like a specific synthesizer's artifacts—rather than universal quality cues. This leads to poor performance when evaluating audio from new, unseen generative models.

To solve this, the researchers employ Domain Adversarial Training (DAT), a technique that forces the model's core feature extractor to learn representations that are useless for predicting the data's source 'domain' (e.g., which dataset it came from), while still being effective for predicting quality. Crucially, they systematically explored how to define these 'domains,' testing strategies from explicit metadata labels to implicit data-driven clustering. Their major insight is that there is no single best definition; the optimal strategy is 'aspect-specific,' varying based on whether the model is evaluating naturalness, intelligibility, or other audio qualities. This tailored approach demonstrably mitigates acoustic bias, resulting in models that correlate better with human judgments and generalize more robustly to novel generative AI scenarios, as evidenced by their acceptance to the IEEE ICME 2026 conference.