Contents
An AI-music detector does not recognize a fake the way a listener notices an uncanny voice. It measures faint statistical traces that a generator, especially the stages that turn a model’s internal representation into a waveform, leaves in the audio, and then maps those traces to a probability at a chosen threshold. It is strong on the systems it was trained on and much weaker on everything else, so understanding what it actually measures is the difference between reading a score well and trusting it blindly.
What a detector actually measures
The signal a detector keys on is a production fingerprint, not a musical judgment. Afchar, Meseguer-Brocal, Akesbi and Hennequin (2025) showed mathematically that the deconvolution modules common to music generators produce systematic peaks at fixed frequencies, a pattern set by the model’s architecture rather than by its training data or learned weights. The effect is precise enough that a roughly 10,000-parameter model reading just those peaks matches a 1.6-million-parameter neural network. In other words, the thing most detectors rely on is a narrow, architecture-specific artifact sitting at particular frequencies, which is why the field is really a set of methods that differ in where they look for it.
The main families of detector
“AI music detector” is not one method but several, and they feed the classifier very different evidence, which is a large part of why two of them can split on the same song (why do music detectors disagree):
- Full-mix artifact detectors. The deployed norm scores the whole track and looks for the generator’s spectral-peak fingerprint. It is cheap and, on a known generator, near-perfect, but it lives or dies on that fingerprint (Afchar et al., 2025).
- Self-supervised front-ends. The strongest research systems start from a large speech model pretrained on ordinary audio, such as wav2vec 2.0 or XLS-R, and fine-tune it with a graph-attention back-end. This is the approach behind the leading singing-deepfake results, and its learned representations generalize across generators better than handcrafted spectral features (Zang, Zhang, Heydari, Duan, ICASSP 2024).
- Music-aware models. SingGraph adds a music-understanding model (MERT) for pitch and rhythm alongside the speech model for lyrics, so it reads musical structure rather than speech-shared artifacts alone. Chen, Wu, Jang and Lee (2024) report relative equal-error-rate improvements of 13.2% on seen singers, 24.3% on unseen singers, and 37.1% on unseen codecs over a strong baseline.
- High-resolution front-ends. Most detectors inherit a 16 kHz speech pipeline, which throws away everything above 8 kHz. Chen and colleagues (2026) argue that this discards high-frequency detail specific to singing, breath and brilliance, and build a 44.1 kHz model that keeps the full band up to 22 kHz precisely to recover those cues.
- Lyrics-transcript detectors. A different regime ignores the signal fingerprint entirely: transcribe the song with speech recognition and classify the words. Because lyrics survive processing that destroys spectral artifacts, this method is far more robust to audio manipulation and to unseen generators (Frohmann, Epure, Meseguer-Brocal et al., 2025).
From a waveform to a score
Whatever the family, the pipeline is the same shape: a front-end turns the audio into features, a classifier scores how strongly those features resemble the synthetic examples it was trained on, and a threshold converts that score into a label. Research usually reports the behavior as an equal error rate because it is threshold-neutral, a single figure describing the whole benchmark. A deployed tool has to pick an actual operating point, and there the two kinds of error diverge: a threshold set to catch more fakes will also flag more real songs. A raw percentage is a position on that curve, not a probability that the track is AI.
Why lab accuracy does not survive the wild
The sub-1% and 99%-plus figures are real and also fragile, because they are measured on the generators the detector learned. Move off that distribution and detection collapses. The same simple detector that scores 100% on the generator versions it trained on drops to 39.83% on an unseen version, below chance (Afchar et al., 2025). The SingFake system rises from a 4.62% equal error rate on seen singers to 42.77% on an unseen language or musical style (Zang et al., 2024), and off-the-shelf speech anti-spoofing models collapse to roughly 50%, no better than a coin flip, when pointed at singing at all. A detector that has never heard your generator is close to guessing, and the whole reason the music-aware and lyrics-based families exist is to fail less badly at that edge.
What this means for reading a result
A music detector is a narrow, capable reader of the fingerprints it was taught and a weak one everywhere else. Treat a flag as a strong first-pass signal when the track is a clean file from a generator the tool likely covers, and as weak evidence when the song is compressed, pitch-shifted, or from an unfamiliar model. The reliable way to use one is to test the original file rather than a re-encode, run more than one method and read agreement as confidence rather than proof, and look for provenance where it exists. For the single-song question this feeds into, see is this song AI-generated.
Sources
- Afchar, Meseguer-Brocal, Akesbi, Hennequin (2025). A Fourier Explanation of AI-Music Artifacts. ISMIR 2025.
- Zang, Zhang, Heydari, Duan (2024). SingFake: Singing Voice Deepfake Detection. ICASSP 2024.
- Chen, Wu, Jang, Lee (2024). Singing Voice Graph Modeling for SingFake Detection. Interspeech 2024.
- Chen, Hu, Huang, Wu, Lee, Jang (2026). Joint Fullband-Subband Modeling for High-Resolution SingFake Detection.
- Frohmann, Epure, Meseguer-Brocal, Schedl, Hennequin (2025). AI-Generated Song Detection via Lyrics Transcripts. ISMIR 2025.