AI vocals in a song: how to spot them?
AI-generated music is no longer just 100% artificial tracks. More and more titles are hybrid: a musician composes and records the instrumental, then generates the vocals; or the other way round, a singer records over an instrumental generated in a few seconds. These tracks are the hardest to spot, both by ear and with a standard detector.
Clues of a generated sung voice
- Diction: soft consonants, blended syllables, distorted words, especially at the end of phrases.
- Breath: breathing that is missing, or added in unnatural places.
- Pitch that is too perfect… or strange: regular pitch slides, like an Auto-Tune that was not set by a human.
- An unstable timbre: the voice seems to change age or person between two sections.
- Backing vocals: harmonies that appear with no arrangement logic.
Why hybrids fool detectors
A detector that analyses the track as a whole makes a kind of average. When half of the sound is human, the overall score drops. In our tests, hybrid tracks scored only 12 to 25% on the full track. They would have passed as “human” without a finer analysis.
The solution: separate before analysing
Our tool first separates the vocals and the instrumental with a source separation model, then analyses each part on its own:
| Track tested | Full track | Instrumental only | Vocals only |
|---|---|---|---|
| Human voice + AI instrumental | 24% | 49% | 7% |
| AI voice + human instrumental | 24% | 5% | 64% |
The first case is the most reliable: the verdict says so (“Instrumental likely AI, possible hybrid track”). The second is much less so: in our tests on real sung voices, the isolated vocals produced too many false alarms. They are therefore shown for information, without being used for the verdict.
What about the spoken voice?
Voice-overs, audiobooks, podcasts, ads: synthetic and cloned voices raise similar questions. It is a different problem from the sung voice: of eight public detectors we tested, six did barely better than chance. Our “Spoken voice” module, in beta, catches 6 out of 10 ElevenLabs voices in our test bench, with 2% false alarms (details). Choose “Spoken voice” on the analysis page.