ARTFEED — Contemporary Art Intelligence

SongBench: New Benchmark for Multi-Aspect Song Quality Assessment

ai-technology · 2026-04-30

Researchers have introduced SongBench, a framework for evaluating text-to-song generation across seven dimensions: Vocal, Instrument, Melody, Structure, Arrangement, Mixing, and Musicality. An expert-annotated database of 11,717 samples from state-of-the-art models was created, labeled by music professionals. Experiments show SongBench correlates highly with expert ratings, revealing performance gaps to guide development toward more professional song generation.

Key facts

  • SongBench evaluates song quality across seven dimensions: Vocal, Instrument, Melody, Structure, Arrangement, Mixing, and Musicality.
  • The database contains 11,717 samples from state-of-the-art models.
  • Samples were labeled by music professionals.
  • SongBench achieves high correlation with expert ratings.
  • The benchmark is designed to diagnose fine-grained performance gaps.
  • It aims to steer development toward more professional and musically coherent song generation.
  • The work addresses limitations in existing evaluation benchmarks for text-to-song generation.
  • SongBench is a specialized framework for fine-grained song assessment.

Entities

Institutions

  • arXiv

Sources