Contents
They are two different instruments pointed at two different parts of the track, and the names get used interchangeably in ways that produce wrong conclusions. An AI music detector reads the full mix for the production fingerprint a generator like Suno stamps across everything it renders. An AI singing-voice detector reads the vocal stem for evidence that the singer is cloned, converted or synthesized. A track can pass one and fail the other for entirely legitimate reasons, so knowing which tool you ran is part of knowing what your result means. Many results that look like contradictions are just the two tools listening to different parts of the song.
The music detector: a family of full-mix readers
An AI music detector is not one method but a family, and they read different parts of the track. The oldest reads a codec fingerprint: deconvolution stages in generators built on residual-vector-quantization codecs stamp systematic spectral peaks at fixed frequencies, a tell so strong that a 10,000-parameter classifier matches a 1.6-million-parameter network on seen generators (Afchar, Meseguer-Brocal, Akesbi, Hennequin, 2025). Newer families read something else: whole-song structure (SONICS, whose long-context model gains F1 from 120 seconds of context but drops 20 to 64 percent F1 across unseen partitions; Rahman et al., 2025), a forensic codec residual (ArtifactNet, F1 0.9993 on the SONICS benchmark; Oh, 2026), or the statistical density of real music (MusicDET, a zero-shot 4.51 percent equal error rate; Han, Wang, Gui, 2026). What they share is that none of them reads a voice. They ask whether the whole rendering came out of a generator.
They also share a fragility. The fingerprint family falls from about 99.8 percent on clean audio to 73.8 percent after 64 kbps MP3, 58.4 percent after 64 kbps AAC, and 66.6 percent after a two-semitone pitch shift (Afchar, Meseguer-Brocal, Hennequin, 2024), and independent testing found a commercial detector “easily fooled by simply resampling audio to 22.05 kHz” (Cros Vila, Sturm, Casini, Dalmazzo, 2025).
The deepest split is which generator built the track
The most important limit on the music-detector side is structural, not a question of accuracy. The codec-fingerprint reader only works on generators whose output passes through a residual-vector-quantization codec, such as MusicGen or the models inferred behind Suno and Udio. Generators built on continuous latent diffusion, such as Stable Audio, AudioLDM and DiffRhythm, have no such codec stage, so they leave no fingerprint comb at all, and the artifact reader is blind to them by construction rather than merely undertrained. The durable readers are therefore the ones that do not depend on the codec: song structure, real-music density, and lyrics. A “not AI” from an artifact detector can simply mean the track was made by a generator that detector cannot physically see.
The singing-voice detector: a vocal-stem examiner
The SVDD family, for singing-voice deepfake detection, asks a narrower question: is the singer in this vocal real, or a clone? The founding benchmark, SingFake, showed both the promise and the ladder of failure: a 4.62 percent equal error rate on seen singers rises to 8.23 percent on unseen singers, 13.62 percent over an unseen codec, and 42.77 percent on an unseen language (Zang, Zhang, Heydari, Duan, 2024). The same work found that off-the-shelf speech anti-spoofing systems land near 50 percent on singing, chance level, which is worth pausing on: the tools built for spoken deepfakes do not transfer to singing either. Singing has accompaniment, wider pitch range, sustained vowels and heavy production, and each difference breaks an assumption a speech model makes.
The dedicated challenge confirmed the pattern. The 2024 SVDD winner reached a 1.65 percent pooled equal error rate, and every system in the field, winner included, scored roughly 50 percent on the unseen commercial generator ACESinger (Zhang, Zang, Shi, Yamamoto, Toda, 2024). Front-end choice matters too: on the controlled CtrSVDD benchmark, raw-waveform features reached 13.75 percent where Mel, spectrogram and MFCC front-ends sat at 25 to 27 percent (Zang, Shi, Zhang, Yamamoto, Han, 2024).
Singing evidence lives where music tools do not look
The clearest technical argument that these are different problems is spectral. Most detectors inherit 16 kHz processing from speech, which discards everything above 8 kHz in the signal. The first systematic high-resolution study of singing-deepfake detection, Sing-HiResNet, works at 44.1 kHz and reports 1.58 percent and 7.45 percent equal error rates on its two WildSVDD test sets, with the discriminative singing information concentrated in the 8 to 16.5 kHz band that speech-rate detectors throw away; above 16.5 kHz the content is less informative (Chen et al., 2026). Singing-aware modeling helps on other axes as well: SingGraph, which reads pitch and rhythm through MERT and lyric content through wav2vec2 XLS-R and fuses them in a graph model, improves 13 to 37 percent relative over a strong speech-derived baseline, reporting 4.01, 6.23 and 6.30 percent equal error rates across its tracks, yet still fails on an unseen language (Chen, Wu, Jang, Lee, 2024). Awareness of singing buys accuracy inside the training distribution. It does not buy generalization.
Lyrics detectors are a third thing
A third family reads neither the vocal stem’s spectrum nor the full mix’s fingerprint. It transcribes the song and classifies the words, so it does not depend on the signal-level artifact at all. Frohmann and colleagues report that a lyrics-transcript detector holds around 85 to 90 percent on an unseen generator where the artifact network collapses to 56.9 percent, and it stays stable under pitch, EQ, noise and reverb (Frohmann et al., 2025). That makes it a useful cross-check rather than a replacement: it can miss cases where the words are human-written or deliberately altered, and it says nothing about whose voice is singing. It answers a content question, while the vocal-stem and full-mix families answer signal questions.
Which tool answers your question
Match the tool to the claim you care about. A wholly generated track, prompt to render, is the music detector’s case, with the caveats in is this song AI-generated?. A cloned vocal is the singing-voice detector’s case, and the hardest version, a cloned vocal over a real instrumental, currently defeats both families, as can a detector catch an AI voice-cover? sets out. Naming the specific generator is a third task again, with the closed-set limits in which AI made this song?. If two tools disagree on one file, the first explanation to check is that they were reading different evidence, the theme of why music detectors disagree. For a single uploaded song the safest workflow is layered: preserve the original, check the whole mix, inspect the vocal stem, compare families, and treat agreement across readers as stronger than any one confident score.
Sources
- Afchar, Meseguer-Brocal, Akesbi, Hennequin (2025). A Fourier Explanation of AI-Music Artifacts. ISMIR 2025.
- Afchar, Meseguer-Brocal, Hennequin (2024). Detecting Music Deepfakes Is Easy but Actually Hard. arXiv:2405.04181.
- Rahman, Paul, Hakim, Sarker, Fattah (2025). SONICS: Synthetic Or Not - Identifying Counterfeit Songs. ICLR 2025. arXiv:2408.14080.
- Oh (2026). ArtifactNet: Forensic Codec-Residual Detection of AI-Generated Music. arXiv:2604.16254.
- Han, Wang, Gui (2026). MusicDET: Zero-Shot AI-Music Detection via Frequency-Guided Normalizing Flows. arXiv:2605.18072.
- Zang, Zhang, Heydari, Duan (2024). SingFake: Singing Voice Deepfake Detection. ICASSP 2024. arXiv:2309.07525.
- Zhang, Zang, Shi, Yamamoto, Toda (2024). SVDD 2024: The Inaugural Singing Voice Deepfake Detection Challenge. arXiv:2408.16132.
- Zang, Y., Shi, J., Zhang, Y., Yamamoto, R., Han, J. (2024). CtrSVDD: A Benchmark Dataset and Baseline Analysis for Controlled Singing Voice Deepfake Detection. Interspeech 2024. arXiv:2406.02438.
- Chen, X., Wu, H., Jang, J.-S. R., Lee, H.-Y. (2024). Singing Voice Graph Modeling for SingFake Detection. Interspeech 2024. arXiv:2406.03111.
- Chen, X., Hu, C.-Y., Huang, S.-F., Wu, H., Lee, H.-y. (2026). Joint Fullband-Subband Modeling for High-Resolution SingFake Detection. arXiv:2604.04841.
- Frohmann, Epure, Meseguer-Brocal, Schedl, Hennequin (2025). AI-Generated Song Detection via Lyrics Transcripts. ISMIR 2025.
- Cros Vila, Sturm, Casini, Dalmazzo (2025). The AI Music Arms Race: On the Detection of AI-Generated Music. TISMIR 8(1).