Automatic singing voice quality assessment is essential for benchmarking modern singing voice generation and conversion systems, yet subjective listening tests are costly and difficult to scale. This paper presents LMM-Net, a label-masked multi-task network that predicts multi-dimensional perceptual scores under incomplete supervision and emphasizes system-level ranking consistency. LMM-Net combines a self-supervised acoustic encoder with an explicit temporal pitch-contour branch. A gated fusion module injects contour cues in a reliability-aware manner, and attention pooling aggregates frame-level evidence into utterance-level representations for multi-task regression. Experiments on the SingMOS-Pro benchmark under the standardized 16 kHz protocol show that LMM-Net improves both utterance-level prediction and system-level ranking. The proposed model achieves a system-level LCC of 0.89 and a system-level SRCC of 0.87 with a system-level RMSE of 0.08, outperforming strong self-supervised baselines augmented with pitch statistics. Ablation studies further confirm that label masking, temporal-contour modeling, gated fusion, and end-to-end adaptation are all necessary to obtain robust gains, especially for system ranking. These results suggest that combining incomplete-label learning with structured temporal melodic modeling is a practical and effective direction for perceptual assessment of singing voice systems.
Copyrights © 2026