APEX: Aesthetic-Informed Popularity Prediction for AI-Generated Music

Source: HuggingFace Daily Papers (2026-05-07) Paper: arXiv 2605.03395 · HF Raw: raw

TL;DR

APEX is a multi-task framework for AI-generated music popularity, trained on 211k songs / 10k hours from Suno and Udio. It jointly predicts engagement-based popularity (streams, likes) and five perceptual aesthetic quality dimensions from frozen MERT audio embeddings. Aesthetic features improve preference prediction on Music Arena (out-of-distribution pairwise human-preference battles across eleven generative music systems), demonstrating that aesthetic and engagement signals capture complementary aspects of AI-generated music.

Tier note

Tier 4. Listed for completeness. The interesting external signal is that AI-generated music is now a measurable consumption surface large enough to support 211k-song training corpora. That is a market-data point worth filing for the industry-pulse thread.