How Much AI Is in This Track? Quantifying the Proportion of AI-Generated Stems in Hybrid Music Mixtures

2026-08-07

AI Summary

This study addresses the limitation of existing AI music detection methods, which are confined to binary classification between fully AI-generated or entirely human-composed audio and cannot quantify the actual proportion of AI involvement in mixed tracks. To overcome this, the work reframes the task as a continuous regression problem by estimating the AI energy ratio α ∈ [0,1] as a measure of AI contribution in hybrid music. The authors construct a multitrack dataset with known mixing proportions of human and AI-generated stems based on neural audio codec reconstructions and propose a CNN architecture capable of both binary classification and α regression. Experiments demonstrate that the regression model achieves a mean absolute error of 0.076 and an R² of 0.85 on held-out test data, while also revealing instrument-specific sensitivities to AI detection and their correlation with spectral characteristics, thereby validating the approach’s applicability in real-world music production scenarios.

Abstract

AI-generated music is increasingly used at the stem level, with producers integrating synthetic drums, basslines, or vocals alongside human-performed instruments. However, current AI music detection systems are binary, treating tracks as either fully AI or fully human. In this paper, we reformulate AI music detection as a regression problem on a continuous AI energy ratio, alpha in [0, 1]. We propose a methodology that leverages a multi-track music dataset to assemble mixtures of human-performed and AI-reconstructed stems (obtained using a neural audio codec) with known proportions of each content type. Using this approach, we first show that a CNN-based model trained on fully AI-generated or human-performed tracks, which achieves >99% accuracy as a binary detector, when faced with mixed content, yields an output that rises with the AI stems' energy contribution, acting as a noisy and miscalibrated estimator. Our analysis of the influence of different stems shows that detection sensitivity depends on the instrument and reflects its frequency content: drums and guitar carry strong codec-artifact signatures, while vocals and bass are less detectable. Based on these insights, we train a similar CNN-based model for regression of alpha, achieving MAE = 0.076 and R^2 = 0.85 on held-out mixtures from the same pipeline. These results suggest that the regression formulation is an initial promising step towards AI-music detection in realistic music production workflows.

Problem

• AI-generated music
• music mixture
• AI detection
• stem-level analysis
• AI proportion quantification

Innovation

• AI music detection
• regression formulation
• neural audio codec
• stem-level analysis
• AI energy ratio