Cross-Cultural Music Similarity: Bridging Human Perception, Signal Processing, and Foundation Models

Understanding similarity across cultural boundaries remains a fundamental challenge in social sciences and artificial intelligence, with music serving as an ideal paradigm for cross‑cultural perception research. This work presents the first comprehensive evaluation of computational music similarity methods against human cross‑cultural music perception, spanning nine diverse musical traditions and encompassing both interpretable signal‑processing features and state‑of‑the‑art foundation models. We collected human annotations from 125 participants across diverse backgrounds, evaluating 1,130 unique audio pairs from Western, Mediterranean, Middle Eastern, Indian, and Chinese musical cultures. Each pair was assessed along three dimensions: overall musical similarity, cultural similarity, and recommendation‑level similarity. We systematically compare these human judgments against computational approaches, including signal‑processing features (rhythm, melody, harmony, timbre) and seven foundation models (MERT, CultureMERT, CLAP, Qwen2‑Audio variants). Results demonstrate that foundation models generally outperform signal‑processing features, with CLAP‑Music&Speech achieving the highest alignment with human perception. Among signal‑processing features, melody consistently emerges as the most predictive of human similarity judgments. Cross‑cultural discrimination analysis reveals that, while humans show strong cultural awareness, computational methods achieve more modest discrimination, with Qwen2‑Audio demonstrating the best cultural boundary detection. Most significantly, ensemble methods combining both approaches achieve substantial improvements, demonstrating the complementary value of interpretable features and learned representations. This study establishes a comprehensive evaluation framework for cross‑cultural music similarity and provides crucial insights for developing culturally aware music technology systems.