Music Generation Evaluation

The field asks whether generated music sounds good, preserves musical structure, and follows instructions about attributes such as genre, rhythm, harmony, and emotion. Approaches combine expert and listener studies with acoustic or symbolic diagnostics, embedding-based similarity metrics, and learned quality or preference models. Benchmarks increasingly test specific capabilities and failure modes, including counterfactual prompt following, multi-track synchronization, song-level quality, and robustness to changed prompts or audio. A central challenge is ensuring that automatic scores reflect human judgments and musical meaning across genres and settings, rather than rewarding superficial similarity or broad quality alone.