How AI Music Detection Works, Signal by Signal

The pipeline

Detection begins with decoding: the compressed file becomes raw sample data. This step already matters, because whatever the encoder threw away is gone for good and cannot be measured later.

Next comes segmentation. Rather than analysing one blob, a detector samples several windows spread across the track — here, up to five. Sampling multiple sections buys something important: the ability to check whether the sections agree with each other, which is itself one of the strongest available signals.

Each segment is then transformed into the frequency domain with a Hann-windowed FFT. The window function matters more than it sounds; chopping a signal abruptly creates spectral splatter that looks like real high-frequency content, which would corrupt exactly the measurement most detectors lean on. From the resulting spectra, a handful of features are derived, weighted and combined into a single estimate.

The six measurements

Each feature is a proxy for a production characteristic rather than a direct measure of 'AI-ness'. Understanding what each one really tracks is the fastest route to understanding why detectors fail.

• Spectral ceiling — the frequency above which energy collapses. Generated audio has often stopped abruptly at a model-imposed limit; so, unfortunately, has any lossy-encoded file.
• Cross-segment spectral distance — how similar the verse and chorus are tonally. Performances drift; single-pass generations often do not.
• Crest factor — peak level against RMS level, a proxy for surviving dynamic range. Low in generated audio, and equally low in any aggressively limited master.
• Spectral-centroid variability — how much the brightness of the track moves over time. Performance produces movement; static synthesis produces less.
• High-band energy ratio — how much of the total energy sits in the top octaves, which reflects both source material and encoding.
• Stereo correlation — whether the image looks captured or constructed. Very high correlation suggests a synthetic or mono-derived image.

Turning features into a probability

Each measurement is mapped onto a score, the scores are weighted, and the weighted sum is converted into a percentage. Two deliberate distortions are then applied. The estimate is shrunk toward 50%, and it is capped between 15% and 85%.

Both are admissions rather than limitations. Shrinkage encodes the fact that these features are weak individually; without it, three mildly suspicious measurements would compound into an unwarranted 95%. The cap encodes the fact that the engine has not been calibrated against a labelled dataset, so any extreme number would be a claim about precision that no evaluation supports.

A separate confidence value is derived mainly from how much the segments agreed and how much usable bandwidth survived encoding. When confidence is low the result is reported as Inconclusive regardless of where the probability landed — an outcome most commercial detectors avoid offering, because 'I don't know' does not demo well.

Signal analysis versus trained classifiers

The detector described here is a deterministic acoustic analysis engine, not a neural classifier. That is a real trade-off and worth stating plainly. A trained model can learn subtle artefact patterns no hand-written rule would catch, and a well-trained one will outperform signal analysis on the generators it was trained against.

The engine's advantages are transparency and stability. Every number is inspectable, the reasoning is reproducible, the same input always yields the same output, and there is no silent degradation as the training distribution ages. You can see precisely why a track scored as it did.

The trained model's disadvantage is generalisation. Accuracy figures come from held-out sets drawn from the same generators the model saw during training, and performance against a generator released after training is unmeasured and typically much worse. An impressive published accuracy figure often describes a world that no longer exists.

Where it all breaks down

Lossy encoding is the biggest problem. A 128 kbps MP3 has had its top end removed by the encoder, which corrupts the spectral ceiling, the high-band ratio and part of the centroid measurement simultaneously. Almost anything sounds machine-made after enough compression.

Aggressive mastering is the second. Loudness maximisation destroys crest factor and flattens dynamics — precisely the fingerprint the detector associates with generated audio. This is the single largest source of false positives, and it hits electronic and pop production hardest.

Editing is the third. Splicing, re-recording, adding live overdubs and running material through analogue-modelled processing all break cross-segment consistency and reintroduce the variability that detection depends on. Human post-production is an extremely effective, entirely unintentional evasion technique.

And short clips are simply insufficient. Under about ten seconds there are too few segments to compare, so the strongest signal available is unavailable.

Why no accuracy number is published here

A credible accuracy claim requires a documented held-out evaluation: a dataset that states its composition, generator versions, encoding conditions and human-music baseline, with results reported as a confusion matrix rather than a single percentage. That work has not been completed for this engine, so no number is published.

Bear in mind what a single accuracy figure conceals even when it exists. A detector that is 95% accurate on a balanced set may sit at a false-positive rate that is unacceptable when most uploads are human — flagging thousands of legitimate tracks at scale. Base rates matter more than headline accuracy, and headline accuracy is what gets advertised.

Until an evaluation exists, treat every result from every acoustic detector — this one included — as a weak signal to be combined with provenance, context and, where it matters, a direct conversation with the artist.