Comprehensive evaluation on real-world human recordings

Conduct a comprehensive study of MuLaCover’s generalization to real-world human recordings beyond the small-scale evaluation reported in the paper.

Background

MuLaCover is trained primarily on synthetic data, so the appendix evaluates its generalization to eight human-recorded songs using Chinese and English examples. Although MuLaCover outperforms the evaluated baselines on most metrics, the authors characterize the evaluation as small and explicitly defer a more comprehensive real-world study to future work.