Researchers quantify AI's role in mixed music tracks

A team of researchers led by Fernando Garcia de la Cruz has developed a novel method to detect and quantify AI-generated components in mixed music tracks. Published on arXiv as part of the ISMIR 2026 conference, their paper introduces a regression-based approach to estimate the proportion of AI-generated stems (alpha) in hybrid music mixtures, addressing a critical gap in current AI music detection systems.

The methodology leverages a multi-track dataset where AI-reconstructed stems (using neural audio codecs) are mixed with human-performed stems in known proportions. Their experiments reveal that traditional binary detectors fail to accurately assess mixed-content tracks, often acting as noisy estimators. The researchers trained a CNN-based model that achieved a mean absolute error (MAE) of 0.076 and an R² score of 0.85 on held-out mixtures, demonstrating significant improvements over existing methods. Notably, detection sensitivity varies by instrument type, with drums and guitar showing stronger codec-artifact signatures compared to vocals and bass.