Can a detector catch an AI voice-cover?

Mostly no, and the reason is structural rather than a matter of tuning. An AI voice-cover, a cloned vocal sung over a real instrumental, is a hybrid: most of the track is genuinely human, and only the voice is synthetic. Every widely deployed detector family was built for a different case, the fully generated track, and the published record shows that the hybrid falls between them. Both families return an answer, but neither answers the question you are actually asking, which is whether that vocal is a cloned voice.

Why a voice-cover is a different problem

The detectors covered in how AI music detectors work read a production fingerprint stamped across the whole mix by a generator such as Suno or Udio. A voice-cover has no such whole-mix fingerprint. The instrumental came from a real session, so most of the signal is clean, and the synthetic evidence is confined to one stem. The question splits in two: is some region of this track synthetic, and whose voice is in it? No deployed system composes both answers, which is why a full-mix detector can be pulled toward the real backing and read the track as human. Which family even applies depends on which half is fake: a cloned vocal over a real beat is the singing-voice detector’s problem, while the mirror case, a real vocal over an AI-generated backing, is the music detector’s, and no single tool spans both.

What an artifact detector sees: a region, not a voice

Artifact detectors can, at best, point at where synthesis lives in a track. Afchar, Meseguer-Brocal and Hennequin (2024) propose patch-level localization for exactly this, flagging which spectrogram regions carry generation artifacts. That is genuinely useful, and it is also the ceiling: a localized “this region is synthetic” verdict carries no identity information. The detector cannot say the cloned vocal imitates a specific artist, because nothing in it reads voices. It reads generator residue. And even that reading is fragile on the pure cases: in the first singing-deepfake challenge, every system, including the 1.65 percent equal-error-rate winner, fell to roughly 50 percent, chance level, on the unseen commercial generator ACESinger (Zhang, Zang, Shi, Yamamoto, Toda, 2024).

What a fingerprint matcher sees: the song, not the singer

The industry’s enforcement layer runs on recording and composition fingerprints, and those identify the work. A matcher such as Audible Magic recognizes a vocal-clone cover through a composition-level match, on the chords, melody, structure and lyrics being the same, not because it knows whose voice was cloned. That is the concession at the center of the hybrid problem: a fingerprint match on a voice-cover tells a rights holder the composition is in use, but it does not tell anyone the vocal is synthetic, or whose it is. The singing-voice research map files this exact case, real instrumental plus cloned vocal, as an open gap that neither commercial family closes.

Singer-ID reads the mix, and attribution has not been tested here

The third family you might reach for, singer identification, degrades on precisely this case. Desblancs, Meseguer-Brocal, Hennequin and Moussallam (ISMIR 2024) showed that singer-ID models operating on the full mix lean on the accompaniment, so on a voice-cover a full-mix model is substantially scoring the real instrumental rather than the cloned voice, which is the one part that matters. Generator attribution does not rescue the situation either: the strongest singing-specific attribution result, 82.37 percent accuracy with LanguageBind, was measured on clean vocals with no accompaniment, in a closed set of eight seen generators labeled A01 to A08 (Phukan et al., 2025). Accompanied, in-the-wild vocals are untested in that benchmark.

The research world knows, and is building the missing benchmark

The gap is now explicitly targeted. The Echoes dataset was built because existing detection benchmarks are dominated by fully generated tracks: it contains 3,577 tracks, about 110 hours, from ten generators, semantically aligned so that detectors cannot take shortcuts, and it deliberately includes the synthetic-vocal-over-real-accompaniment case (Pascu, Oneata, Cucu, Müller, 2026). That a 2026 dataset had to be constructed to test this scenario is itself the honest status report: the hybrid case is where evaluation is only beginning, not where reliable products are.

What you can still do with a suspected voice-cover

Treat the tools as partial witnesses, and treat a detector score as a triage signal rather than a verdict. A full-mix music detector, as covered in is this song AI-generated?, answers the whole-track question and can be wrong in both directions on a hybrid; independent testing found one commercial detector “easily fooled by simply resampling audio to 22.05 kHz” (Cros Vila, Sturm, Casini, Dalmazzo, 2025). A vocal-stem detector, the family described in AI singing-voice detector vs AI music detector, asks the right question about the voice but inherits its own generalization walls. If the real question is which model made the vocal, the closed-set limits in which AI made this song? apply. A high score justifies preserving the original file, separating stems, and comparing the isolated vocal against the alleged singer; a single “not AI” verdict is close to no information, because the tool that produced it was probably not looking at the voice. Run more than one family and weigh agreement as confidence rather than proof.

Sources

• Afchar, Meseguer-Brocal, Hennequin (2024). Detecting Music Deepfakes Is Easy but Actually Hard. arXiv:2405.04181.
• Zhang, Zang, Shi, Yamamoto, Toda (2024). SVDD 2024: The Inaugural Singing Voice Deepfake Detection Challenge. arXiv:2408.16132.
• Desblancs, Meseguer-Brocal, Hennequin, Moussallam (2024). From Real to Cloned Singer Identification. ISMIR 2024.
• Phukan, O. C., Girish, Akhtar, M. M., Behera, S. R., Mallick, P. (2025). Towards Source Attribution of Singing Voice Deepfake with Multimodal Foundation Models. arXiv:2506.03364.
• Pascu, O., Oneata, D., Cucu, H., Müller, N. M. (2026). Echoes: A Semantically-Aligned Music Deepfake Detection Dataset. arXiv:2603.23667.
• Cros Vila, Sturm, Casini, Dalmazzo (2025). The AI Music Arms Race: On the Detection of AI-Generated Music. TISMIR 8(1).