Researchers Estimate AI-Generated Share in Hybrid Music
Researchers posted a study on August 7 that reframes AI-music detection as estimating an AI energy ratio within hybrid tracks instead of making a human-or-AI binary call. On 21,212 controlled mixtures from 240 MoisesDB tracks, a five-second CNN regressor achieved MAE of 0.076 and R-squared of 0.85, but the team did not test commercial generators or normal studio post-processing.
Researchers posted an arXiv paper on August 7 that tests whether an audio model can estimate how much of a hybrid music mix comes from AI-reconstructed stems. The work, listed by arXiv as accepted at ISMIR 2026, replaces the usual human-versus-AI label with a continuous AI energy ratio between zero and one.
From a binary label to a proportion
The team built 21,212 controlled mixtures from 240 professionally produced tracks in MoisesDB. Each track contained between two and 10 isolated stems, such as vocals, drums, bass, guitar and other instruments. For each mixture, selected human-performed stems were passed through Meta's EnCodec neural audio codec at 3 kbps, while the remaining stems stayed in their original form. The target ratio measured the share of acoustic energy contributed by the reconstructed stems.
A CNN trained only to separate fully original from fully reconstructed tracks exceeded 99% accuracy at those endpoints. That strong binary result did not translate into a reliable proportion estimate: using its output as an implicit ratio produced a mean absolute error of 0.289 and an R-squared of -0.85 with five-second windows. A CNN with the same backbone trained directly on the continuous ratio reached a mean absolute error of 0.076, R-squared of 0.85 and Pearson correlation of 0.93 on held-out mixtures from the same pipeline.
The experiment also found uneven sensitivity by instrument. Codec artifacts were easier to detect in drums and guitar, while vocals and bass contributed weaker signals. That variation means a whole-track score can depend on which stem was reconstructed, not only how many stems were changed.
A controlled method, not a production detector
The result does not show that the model can measure the AI share of commercial releases. The experiments used one codec at one bitrate, and the reconstructed stems preserved human-performed musical content rather than coming from commercial generators such as Suno or Udio. The mixtures also excluded common production steps including equalization, compression, level balancing and creative editing.
The practical contribution is therefore the evaluation target and controlled test method. It gives music-information-retrieval researchers a way to study hybrid material more precisely than a binary label, while the paper's limitations make clear that generator diversity, studio processing and instrument-aware detection must be tested before the approach can support real-world provenance decisions.
Key Points
- 1The study reframes AI-music detection as estimating a continuous AI energy ratio in hybrid tracks rather than assigning a binary label.
- 2A five-second CNN regressor reached MAE of 0.076 and R-squared of 0.85 on held-out mixtures, while a binary detector used as a ratio estimator performed poorly.
- 3The controlled setup used one codec and no commercial generators or studio post-processing, so the result is not a deployment-ready detector.
Scoring Rationale
The proportional-detection framing and controlled benchmark are technically useful, but practical impact is limited by single-codec reconstructed stems and the absence of commercial-generator and studio-processing tests.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
