How the AI music detector works
AI-generated music is not “recorded”: it is made by a model that directly produces the audio signal. This process leaves subtle traces in the sound spectrum, inaudible to the ear but detectable by a model trained to recognise them. Here is how the tool works, step by step.
1. Reading the file and its metadata
The file is first read and its metadata inspected: some generators or export tools leave their name in the file information, or a provenance signature (Content Credentials / C2PA). Their absence proves nothing, as they are very easily erased.
2. Analysing the full track
The track is cut into overlapping 5-second windows (one every 2.5 seconds). Each window is turned into a spectrogram and then scored by a detection model. This gives an AI probability for every moment of the track, shown on the report timeline. A second model looks at 2-minute excerpts for an overall view.
3. Vocals / instrumental separation
The track is then split into stems with Demucs, a source separation model developed by Meta's research team: vocals, drums, bass and other instruments. This rebuilds the instrumental only and the vocals only.
4. Analysing each part
The instrumental and the vocals are analysed separately, with the same window method. The instrumental alone reveals hybrid tracks where a real voice sits on an AI-generated instrumental: in our tests, they scored only 12 to 25% on the full track, and stood out clearly once the vocals were removed.
The isolated vocals are shown for information but are not used for the verdict: in our tests on real sung voices, they produce too many false alarms (processed vocals, choirs, effects).
5. The verdict
The verdict keeps the higher of the two reliable readings, the full track and the instrumental only:
| AI score | Conclusion |
|---|---|
| below 20% | Very likely human |
| 20 to 45% | Likely human |
| 45 to 70% | Likely AI |
| above 70% | Very likely AI |
The verdict always gives a trend. When the readings disagree, it says so, for example “Instrumental likely AI, human voice”.
The models used
Detection relies on the SONICS / SpecTTTra models, released openly by researchers in 2024 and trained on a dataset of about 97,000 tracks, nearly half of them generated by Suno and Udio. Separation uses Demucs (htdemucs). The thresholds and the isolated-vocals scale were calibrated on our own tests, detailed on the Reliability page.
What about the spoken voice?
In “Spoken voice” mode, the recording is cut into 15-second passages, each analysed by a speech recognition model (wav2vec2 XLS-R) specialised in detecting synthetic voices, complemented by a layer we trained on more than 2,000 real and synthetic voices in five languages. No stem separation: the model works directly on the voice. The module is in beta; its results are detailed on the Reliability page.