AI audio detection is not the same as AI music detection

Speech detection has advantages music does not

Synthetic-speech detection is a comparatively well-posed problem. There is usually one source, a narrow bandwidth, an unprocessed or lightly processed signal, and a strong prior on what a human vocal tract can physically do. Breath noise, glottal behaviour, formant transitions and prosody all provide structure a detector can exploit.

Music removes almost all of those advantages

• Many overlapping sources. Drums, bass, guitars, synths and vocals occupy the same spectrum. Any per-source artefact is masked by everything else.
• Heavy deliberate processing. Compression, saturation, pitch correction, time-stretching and limiting are normal, wanted parts of human music production — and they look exactly like the “artefacts” a speech detector flags.
• Full bandwidth. Music uses the whole spectrum, so a speech model tuned to telephone or podcast bandwidth is looking in the wrong place.
• Enormous stylistic variance. Lo-fi, black metal, chiptune and ambient are all legitimately “unnatural”. There is no acoustic prior on what music must sound like.
• Post-production is universal. Even generated music is usually mastered after generation, smoothing the very differences a detector needs.

Practical consequences

A speech-focused detector run on a song typically returns high-confidence nonsense — often confidently “AI” for any heavily processed track, because processing looks synthetic to a model trained on clean voice recordings. Confident output on out-of-domain input is the classic failure mode of a well-trained model asked the wrong question.

What a music-specific tool does differently

• Samples multiple sections and checks whether they agree, rather than scoring one clip
• Measures arrangement-level structure, not just frame-level texture
• Treats loudness maximisation as an expected confound rather than a red flag
• Reports lower confidence when encoding has removed the detail it depends on
• Declines to answer when the evidence is contradictory

Our detector does all of these, and documents where it still fails on the accuracy page.

What about a song with an AI voice over human instruments?

That is a third problem again, and the hardest of the three. Doing it properly requires stem separation before analysis — see AI vocal detection and its limitations.

For songs, start with the music-specific free AI music detector rather than a general audio tool.