SongEval is a music evaluation dataset jointly released by Shanghai Conservatory of Music, Northwestern Polytechnical University, University of Surrey and Hong Kong University of Science and Technology in 2025. The related paper results are "SongEval: A Benchmark Dataset for Song Aesthetics Evaluation”, which aims to provide an aesthetic evaluation of complete songs.

The dataset contains more than 2,399 songs (including vocals and accompaniment), and the data was annotated by 16 expert raters on five perceptual dimensions (overall coherence, memorability, naturalness of vocal breathing and phrases, clarity of song structure, and overall musicality). It covers approximately 140 hours of high-quality audio, including Chinese and English songs and 9 mainstream genres.

Dataset structure:

• audio : WAV audio of the complete song.
• gender : Male or female.
• aesthetic_scores : A dictionary of 5 human-annotated scores.