DeRA-MOS boosts text-to-music evaluation with decoupled ranking and alignment

Evaluating text-to-music (TTM) systems typically requires expensive human mean opinion scores (MOS) for both music impression (MI) and text alignment (TA). Existing automatic MOS estimators rely on point-wise regression or distributional classification, which fail to directly optimize rank-based metrics and provide weak geometric constraints for cross-modal coherence. To address these gaps, researchers from Academia Sinica and National Taiwan Normal University introduce DeRA-MOS, a decoupled optimization framework.

DeRA-MOS separates the evaluation into two components. For MI, it employs a batch-aware listwise ranking loss that models relative order within each mini-batch, aligning training with Spearman's rank correlation coefficient (SRCC). For TA, a score-anchored modality alignment loss maps human scores to target audio-text similarity and regularizes the latent space before fusion. Experiments on the MusicEval benchmark demonstrate that DeRA-MOS substantially improves both MI and TA ranking metrics, establishing a robust paradigm for large-scale TTM evaluation. The paper has been accepted to IEEE Signal Processing Letters.