New AI framework MSAF+HiGIA outperforms in song aesthetics evaluation
Two novel modules fuse vocals and accompaniment to judge song quality like a human.
A team of researchers (Yishan Lv, Jing Luo, et al.) has introduced a new framework for automated song aesthetics evaluation, addressing a gap where prior work focused on speech or singing quality rather than holistic song aesthetics. The framework features two key modules: Multi-Stem Attention Fusion (MSAF), which builds bidirectional cross-attention between mixture-vocal and mixture-accompaniment pairs to capture complex musical interplay, and Hierarchical Granularity-Aware Interval Aggregation (HiGIA), which learns multi-granularity score probability distributions and then regresses within an interval for the final score.
Tested on two datasets—SongEval (AI-generated songs) and an internal aesthetics dataset (human-created songs)—the method outperformed two state-of-the-art models on multi-dimensional metrics. Unlike conventional approaches that predict a single Mean Opinion Score (MOS), HiGIA models uncertainty in human perception, producing more nuanced evaluations. The inference code and checkpoint are openly available on GitHub. The paper has been accepted at ISMIR 2026, highlighting its significance for music AI quality assessment.
- Multi-Stem Attention Fusion (MSAF) enables bidirectional cross-attention between vocal and accompaniment stems for richer musical feature extraction.
- Hierarchical Granularity-Aware Interval Aggregation (HiGIA) models multi-granularity score distributions instead of a single MOS, better capturing human perception nuances.
- Outperforms two SOTA models on both AI-generated (SongEval) and human-created song datasets, with public code and checkpoint.
Why It Matters
Automated song aesthetics evaluation becomes more human-like, enabling better quality control for AI-generated music and streaming platforms.