Multifractal Comparison of Billboard and AI-Generated Music

摘要

Musical structure spans nested timescales, from fine-grained fluctuations to long-range organization. To capture this, we apply Detrended Fluctuation Analysis (DFA) and Multifractal Detrended Fluctuation Analysis (MFDFA) to compare human-composed music (Billboard Top 5 hits, 1950-2024) with AI-generated outputs from Suno, DiffRhythm, and YuE. While all models capture fractal properties of music, differences persist. Suno aligns most closely with human music but shows reduced fine-scale variability. DiffRhythm yields narrower spectra and lower complexity, while YuE matches large-scale structure yet exhibits greater small-scale variability. Decade-level analysis shows divergence is smallest for earlier, more homogeneous eras (1950s-1960s) and greatest during periods of stylistic diversity and production complexity (2000s-2020s). We propose integrating fractal descriptors into training objectives and refining architectures to improve temporal sensitivity, advancing AI systems toward more structurally representative and authentic music generation. The code and results can be found at https://github.com/zhangkkevin/billboard-ai-fractal-comparison.