Quantifying AI-Generated Stems in Hybrid Music Mixtures
A recent study published on arXiv (2608.07285) introduces a novel technique for measuring the percentage of AI-generated stems in hybrid music compositions, advancing beyond simple binary detection of AI. The researchers redefine AI music identification as a regression challenge focused on a continuous 'AI energy ratio' (alpha in [0,1]). They create blends of human-performed and AI-generated stems (utilizing a neural audio codec) with established ratios. A CNN-based model, which is trained exclusively on either AI or human tracks, achieves over 99% accuracy as a binary classifier. However, when analyzing mixed content, its output increases with the AI stems' energy contribution, serving as an unreliable and poorly calibrated estimator. The study evaluates the implications of this method, proposing a more sophisticated tool for recognizing AI's role in music creation.
Key facts
- Paper on arXiv: 2608.07285
- Proposes continuous AI energy ratio (alpha) for detection
- Uses multi-track dataset to create mixtures
- CNN model achieves >99% binary accuracy
- Binary detector output is noisy and miscalibrated on mixed content
- Method uses neural audio codec for AI-reconstructed stems
- Reformulates AI music detection as regression problem
- Focus on stem-level AI integration in music
Entities
Institutions
- arXiv