Dual-engine AI music detector survives MP3

The clever part isn’t just stacking two models, it’s using a cheap confidence gate so the expensive separation pass only runs when the classifier is unsure. That makes the system feel more production-shaped than a single end-to-end detector, even if the edge cases are still messy.

• –Mel-spectrogram CNNs can look strong on WAV and then lose the signal once lossy compression strips the artifacts they learned.
• –Demucs adds a different hypothesis: human recordings leak across stems, while fully synthetic tracks tend to reconstruct too cleanly after separation and remixing.
• –The compute tradeoff is sensible, because source separation is costly and shouldn’t run on every track if the CNN already has high confidence.
• –The biggest risk is generalization: different generators, mastering chains, and Demucs nondeterminism can all move borderline samples around.