SongEval

dataset·song aesthetics·active

Full-length real and generated songs with expert ratings across five aesthetic dimensions.

Recorded facts

Official site	https://huggingface.co/datasets/ASLP-lab/SongEval ↗
Geography	China
arxiv	2505.10793
github repository	ASLP-lab/SongEval
huggingface repository	ASLP-lab/SongEval
size	songs: 2399 · expert annotators: 16 · rating dimensions: 5 · duration hours approx: 140
scope	Human-aligned aesthetic evaluation of complete generated songs.
genres	9 mainstream genres including pop, rock, jazz, hip-hop and classical
steward	Northwestern Polytechnical University ASLP Lab and collaborators
creators	Jixun Yao; Guobin Ma; Huixin Xue; Huakang Chen; Chunbo Hao; Yuepeng Jiang; Haohe Liu; Ruibin Yuan; Jin Xu; Wei Xue; Hao Liu; Lei Xie
languages	English; Mandarin Chinese
modalities	full-song WAV; five-dimensional ratings; overall scores; gender metadata
geographies	Global
access method	Hugging Face dataset and GitHub evaluation toolkit.
intended uses	song aesthetics prediction; music-generation evaluation; reward modeling
related papers	SongEval: A Benchmark Dataset for Song Aesthetics Evaluation
annotation method	Musically trained annotators used five-point Likert ratings on five perceptual dimensions.
collection method	Outputs from five generation models plus real and bad-case samples selected for annotation.
provenance claims	Dataset card identifies generated and real/bad-case sources but does not fully expose rights provenance per song.
current availability	Dataset and toolkit available.
us market scope basis	Publicly available or materially used in US-facing music/audio-AI research.
limitations biases disputes	Aesthetic judgments are subjective and annotator-cultural coverage is not fully described.; Model outputs can encode generator-specific artifacts.
related models tools benchmarks	ICASSP 2026 Automatic Song Aesthetics Evaluation Challenge; MARBLE