Abstract
The recognition of emotion in singing voices remains a complex frontier in computational paralinguistics, constrained fundamentally by the scarcity of annotated datasets and the profound influence of cultural diversity on musical expression. Traditional supervised learning paradigms demand exhaustive labeling, which is prohibitively expensive and largely restricted to Western musical traditions. To dismantle these barriers, this paper introduces a novel framework termed Culture-Adaptive Contrastive Pretraining. This methodology leverages vast repositories of unannotated global singing data by integrating cultural metadata as a structural prior within a self-supervised contrastive learning objective. By defining positive and negative acoustic pairs through a culturally informed lens, the proposed architecture preserves nuanced, region-specific emotional articulations that standard augmentation techniques frequently eradicate. The framework is systematically evaluated under severe low-label conditions, utilizing only marginal fractions of annotated data for downstream fine-tuning. Comprehensive experimental analyses reveal that the culture-adaptive approach significantly outperforms standard contrastive baselines in cross-cultural generalization and feature robustness. Furthermore, ablation studies delineate the efficacy of culturally neutral augmentations and metadata-guided loss formulations. The findings presented herein not only establish a new benchmark for low-resource singing emotion recognition but also provide a scalable, ethically conscious blueprint for modeling global musical diversity in artificial intelligence systems.