AI-music detector · audio
Was this track generated by AI?
The neural audio codecs inside AI-music generators (Suno, Udio and their kin) leave a faint, regular comb of peaks high up in the spectrum, between about 5 and 16 kHz. It comes from how their decoder upsamples, so it is baked into the architecture, not the song. This averages the spectrum, fits a baseline under it, and measures how far a comb lifts above that baseline. Read it as a strong lead on codec-based generators, not proof.
Runs entirely in your browser. Your file is never uploaded or stored.
This teaches one basic forensic signal, read in isolation. It is a lead, not a verdict, and can be fooled on its own — real confidence comes from combining several independent signals and understanding the relationship between them.
Decoding and analysing the spectrum…
The grey line is the averaged spectrum in the high band; the dim line beneath it is the baseline fitted under the dips. A real recording rides right on its baseline, so the orange fill is almost nothing. A neural-codec generator lifts a run of evenly spaced spikes several dB above the baseline — that orange comb is the fakeprint.
What the comb is
Almost every AI-music generator builds its sound through a neural audio codec or a stack of transposed convolutions. Afchar and colleagues at Deezer proved that these upsampling layers cannot help stamping a regular pattern of small peaks into the high frequencies, the audio cousin of the checkerboard artifact seen in AI images. Their key result is that this comb depends on the generator's architecture, not on the training data or the model's weights, which is why it shows up across open models and commercial services like Suno and Udio alike, and why a simple spectral test matches deep-learning detectors at over 99% accuracy on the generators they studied. The tool averages the spectrum across the whole track, fits a lower hull under it as the baseline, and reads how far the spectrum rises above that hull between 5 and 16 kHz.
How to read the result
A band that sits flat on its baseline, with a peak barely a decibel above it, is what a human recording looks like. A comb of evenly spaced spikes lifting three, four, five or more decibels above the baseline is the generator tell worth acting on. The number to watch is the peak above the baseline: real music tends to stay under about two decibels, AI music pushes well past it. Because the comb is periodic, its shape in the panel is often more convincing than any single threshold.
Where it fails
This signal is powerful but narrow, and the same researchers were careful to map its limits in a companion paper. It reads the fingerprint of the neural codec, so it is at its best on codec-based generators (the Suno and Udio family) and is structurally blind to a whole other class, the continuous-latent diffusion generators such as Stable Audio and AudioLDM, which leave no codec comb at all. It is also fragile: a pitch-shift, added noise, or a low-bitrate MP3 re-encode smears or erases the high-frequency comb, so a shared or re-compressed file can read clean even when it was generated. And it is version-sensitive even within the codec family, where an unseen model revision can drop detection sharply. Treat a strong comb as a solid reason to escalate; treat a clean read as ruling out very little. For the compression side of the same high band, see the codec history tool.
Sources
- Afchar, D., Meseguer-Brocal, G., Akesbi, K. & Hennequin, R. (2025). A Fourier Explanation of AI-Music Artifacts. Proc. 26th Int. Society for Music Information Retrieval Conf. (ISMIR 2025), Daejeon. arXiv:2506.19108.
- Afchar, D., Meseguer-Brocal, G. & Hennequin, R. (2024). Detecting Music Deepfakes Is Easy but Actually Hard. arXiv:2405.04181.