{"id":"85c4dea7-1272-4701-bcbf-e564b2d9a75c","arxiv_id":"2506.19108","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Zero-upsampling in audio generators replicates the low-frequency spectrum at fixed intervals, creating architecture-dependent spectral peaks that enable a simple, interpretable detector for AI-generated music.","lead":"This paper shows why AI-generated music has tiny, evenly spaced frequency peaks: the deconvolution layers that expand compressed audio copy the low-frequency spectrum at regular intervals. The authors derive where those peaks appear, confirm them in open-source and commercial music generators, and use them in a small linear detector that reaches about 99% accuracy.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof establishes periodization after zero-upsampling, but not persistence through the kernel and nonlinearities; observable artifacts are not actually proven.","rationale":"The reader's weakest_assumption precisely identifies the same gap: artifact persistence through the convolution kernel and subsequent nonlinearities is assumed rather than proven. The paper itself flags this in Sections 3.2 and 3.4, so this is not a manufactured objection. The DSP periodization lemma is sound, and the empirical results on open-source models are meaningful, but the abstract's 'mathematically prove' overstates the scope of the proof. Because the central detection claim depends on the artifacts being observable in outputs, and because a concrete cancellation scenario is admitted in the text, the appropriate verdict is CONDITIONAL rather than full acceptance: the authors should either prove a bound on kernel-induced attenuation under realistic training conditions or soften the claim to 'periodization provably occurs; empirical evidence shows it survives in tested models.' I found no additional load-bearing flaw beyond this persistence gap; the closed-source comparison unfairness and the Udio 32 failure are secondary qualifications, not fatal to the core idea.","tokens_in":10378,"tokens_out":2445,"duration_ms":30921,"concrete_test":"Using the released Encodec, DAC, and Musika! checkpoints, compute for each deconvolution layer the DFT of the learned kernel zero-padded to the appropriate signal length, and cascade the squared magnitudes across layers at the predicted peak-frequency grid (n·fs for each stride). Then feed a unit-DC vector through the full layer including ReLU and measure the output spectrum at those frequencies. If any predicted peak is attenuated by more than, say, 20 dB relative to neighboring bins in the end-to-end transfer, that peak's presence cannot be attributed to the proven periodization alone; if all predicted peaks survive, the persistence concern is empirically settled for those models.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that deconvolution outputs exhibit detectable peaks at n·fs rests on an unproven persistence assumption. Section 3.2 proves only that zero-upsampling periodizes the hidden-layer spectrum; the subsequent 1-strided convolution multiplies that spectrum by the kernel's frequency response. The paper explicitly concedes that a constant kernel of size equal to the stride cancels all peaks except DC, and responds only that small kernels are 'unlikely to remove all those artifacts entirely.' That is a heuristic, not a proof. Section 3.4 further admits that nonlinear activations, skip connections, and batch normalization are not covered by the theory. Yet the abstract states that the outputs 'mathematically prove' exhibit systematic frequency artifacts, and the detector's rationale depends on these peaks being reliably present in generated audio. The empirical demonstrations on Encodec, DAC, and Musika! show survival in those specific trained checkpoints, but they do not close the logical gap: a model could in principle learn kernels whose frequency response nulls the predicted peak positions, making the artifact architecture-dependent only in the weak sense that the periodization exists in an intermediate representation, not in the output. This is the load-bearing soft spot because if the persistence assumption fails for any practical generator, the position prediction and the architecture-inherent claim both lose force.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies spectral artifacts in AI-music generation. It shows that the zero-upsampling step of a deconvolution (transposed convolution) with stride k periodizes the hidden-layer spectrum, so that a DC or low-frequency component is replicated at multiples of the input sampling rate, producing predictable spectral peaks. The authors extend this to sequential CNNs with a recursive count of peaks and use the predicted peak positions to build a linear logistic-regression detector on an 'artifact fingerprint.' Experiments on Encodec, DAC, Musika!, Suno, and Udio are reported, with detection accuracy above 99% in several settings. The manuscript claims that the artifacts are architecture-dependent and independent of training data and learned weights.","tokens_in":10579,"tokens_out":9437,"duration_ms":96801,"significance":"The core observation is interesting and timely: if valid, it provides an interpretable, parameter-free explanation of why AI-music detectors work, and it predicts that artifact locations are fixed by stride hyperparameters rather than by training data. The paper is commendably concrete: the code repository is referenced, the peak positions are derived rather than fit, and the detector is evaluated on held-out splits. The strongest contribution is the framework connecting transposed-convolution arithmetic to Fourier periodization, together with the demonstration that a simple linear model can rival black-box detectors in some settings. However, the theoretical claim currently outruns the proof: only the zero-upsampling periodization is proven, while survival through kernels and nonlinearities is treated heuristically, and the closed-source parity claim is based on an unmatched evaluation protocol. These issues are load-bearing for the abstract's central claim, though they are addressable within the manuscript's scope.","major_comments":[{"comment":"The abstract states that the paper 'mathematically prove[s] that their outputs exhibit systematic frequency artifacts,' but the formal derivation in §3.2 establishes periodization only for the zero-upsampling stage of a deconvolution. The subsequent 1-strided convolution multiplies the periodized spectrum by the kernel's frequency response, and the manuscript itself concedes that a constant kernel of size equal to the stride would cancel all of the predicted peaks except DC, answering only that small kernels are 'unlikely to remove all those artifacts entirely.' Section 3.4 explicitly states that persistence through nonlinear activations, skip connections, and batch normalization is not covered by the theory. Since the central claim and the detector's rationale require the peaks to be present in the final output, the authors should either prove a persistence result under precise conditions or explicitly downgrade the claim to a theorem about the intermediate representation plus an empirical persistence hypothesis. This is the main load-bearing gap, and it affects the title-level claim.","section":"Abstract; §3.2; §3.4"},{"comment":"The conclusion that the artifacts are 'solely architecture dependent' goes beyond the evidence presented. Figure 4 shows the same peak positions across four DAC models trained with different seeds and datasets, which is good evidence of position invariance for one architecture family, but it does not establish that artifacts are independent of learned weights in general; the peak amplitudes differ between runs, and only one model configuration is tested. To support the stronger claim, the authors should either verify position invariance on at least one additional architecture under multiple seeds, or phrase the claim as invariance of peak locations rather than of the artifacts themselves.","section":"§4.2, Fig. 4"},{"comment":"The claimed parity with SpecTTTra-α in Table 2 is not supported by a matched evaluation. As the authors note, the SONICS real-audio tracks were not available and FMA tracks resampled to 16 kHz were used instead, so the scores are not directly comparable with the numbers reported in [16]. The abstract's statement that the method is 'on par with deep learning-based approaches' and the conclusion's 'on par with previous million-parameter models' therefore overstate the evidence for the closed-source scenario. The authors should either run the detector under the original SONICS protocol or restrict the parity claim to the open-source comparison in Table 1, where the baseline was evaluated under the same protocol.","section":"§4.3, Table 2"},{"comment":"The open-source detection experiment uses auto-encoded real FMA audio as the 'synthetic' class rather than outputs of an end-to-end generative model. This is a reasonable codec-artifact study, and it matches the protocol of [17], but the abstract and introduction frame the result as AI-music detection. The gap between codec reconstruction detection and generated-music detection should be acknowledged explicitly in the claims, even though the Suno and Udio experiments partially bridge it, because the transfer from codec artifacts to full generators is itself an assumption rather than a proven fact.","section":"§4.3, Table 1"}],"minor_comments":[{"comment":"The sentence 's and v have exactly the sample spectrum' appears to mean 'same spectrum'; please fix the typo and clarify that the equality holds for the continuous-time Fourier transforms, while the discrete-time spectra differ because of the change in sampling rate.","section":"§3.2"},{"comment":"The notation in the recursive count is hard to follow: layer indices appear both as subscripts and superscripts, and the claim that a spectrum is 'cloned k(i+1)/2 times ... creating k(i+1) half-spectra' should be expanded with a two-layer worked example so that the product formula P_max is transparent.","section":"§3.3"},{"comment":"The phrase 'successive 213 samples' should be typeset as 2^13 samples or explained in words; as printed it is ambiguous.","section":"Fig. 3 caption"},{"comment":"The artifact fingerprint is defined as the average spectrum minus a sliding-window local minimum; the window size and bandwidth choices should be given numerical values, since the detector's behavior and the interpretation as a 'fingerprint' depend on them.","section":"§4.2"},{"comment":"The dagger symbols for versions unseen during training are defined only in the text after the table; please move the definition into the table caption or a footnote so that the table is self-contained.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a valuable central observation and a working detector, but the mismatch between the proof and the advertised claim is real. I recommend major revision rather than rejection because the core periodization result is correct and the empirical evidence is substantial; the authors can fix the paper by narrowing the theoretical claim and aligning the evaluation claims with the actual protocol. I would also ask the editor to ensure the authors address the SONICS comparison, since the current unmatched protocol risks misleading readers about state-of-the-art parity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper gives the clearest account yet of why transposed-convolution generators leave spectral peaks, and a 10K-parameter logistic regression that exploits them is genuinely competitive on open-source codecs. But the abstract's 'mathematically prove' is stronger than what Section 3 establishes: zero-insertion periodization is proven, survival through the convolution kernel and nonlinearities is argued heuristically and then demonstrated empirically. That distinction matters, though it doesn't sink the paper.\n\nWhat's new: the recursive multi-layer peak-count formula (product of strides) that predicts, say, 161 peaks for Encodec's {8,5,4,2} strides; the demonstration that peak positions are stable across training seeds and datasets; and the framing of detector failure modes (resampling, pitch shift) that follows from the theory. The Fourier derivation itself is textbook — spectral imaging from zero-upsampling is classical — and they cite the prior checkerboard/audio artifact literature. The credit goes to making the prediction concrete and testing it on real models, with a public code repo.\n\nWhere it's soft: the persistence assumption. Section 3.2 proves zero-upsampling periodizes the spectrum, but the 1-strided convolution multiplies by the kernel's frequency response; a constant kernel of the right size would null all non-DC replicas. The paper acknowledges this and says small kernels are 'unlikely' to remove all artifacts — that's a heuristic, not a proof. Section 3.4 extends the caveat to nonlinearities, skip connections, batch norm. The empirical evidence on Encodec, DAC, Musika! shows the peaks actually survive in those trained checkpoints, but it doesn't close the logical gap. The closed-source comparison is also explicitly unfair (real tracks not released), so the 'on par' with SpecTTTra there is weaker than the Table 1 results. The Udio 32 zero-shot failure is a real bound.\n\nBottom line: this is a useful, mechanically grounded paper that deserves a serious referee. I'd recommend accepting it for review with a request to soften the proof language in the abstract and Section 3.2, and to move the persistence claim from 'proven' to 'predicted and empirically confirmed on the studied architectures.' The core contribution — architecture-dependent, data/weight-independent peak positions — holds up for the open models, and the detector is a nice byproduct. I'd bring it to reading group and cite it if I work in that area.","headline":"Solid mechanistic account of transposed-convolution spectral artifacts, but the abstract's 'mathematically prove' overstates what is actually proven.","tokens_in":11132,"tokens_out":2266,"would_cite":true,"duration_ms":20792,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Fourier proof shows AI music is detectable because deconvolution layers stamp predictable spectral peaks into every generated clip.","keywords":["AI music detection","deconvolution artifacts","Fourier analysis","spectral peaks","checkerboard artifact","architecture-dependent artifacts","interpretable detection","generative audio"],"falsifier":"Construct or train an audio generator whose deconvolution kernels are constant over the stride length, so their frequency response cancels the replicated peaks, while hidden layers are kept free of a DC bias; if the predicted peaks at $n f_s$ still appear in the average spectrum, the theory as stated is incomplete, and if they vanish, the artifact is not strictly unavoidable.","tokens_in":10118,"feed_emoji":"🎵","tokens_out":5391,"duration_ms":52212,"temperature":0.7,"pith_summary":"This paper tries to establish that the upsampling layers inside neural audio generators leave a permanent, detectable signature in the frequency spectrum of the music they produce: a row of small peaks whose positions are fixed by the architecture's stride sizes and do not depend on training data or learned weights. The argument is a Fourier-domain proof: zero-insertion upsampling periodizes a layer's spectrum, and if hidden layers carry a strong DC or low-frequency component from biases and ReLU activations, that component is cloned at predictable frequencies. The authors then show the pattern recurs through stacked deconvolution layers, producing a fractal-like set of peaks, and that a simple linear classifier reading this artifact fingerprint can match deep-learning detectors, exceeding 99% accuracy on several tests. If true, this gives an interpretable, architecture-level explanation for why synthetic music is detectable and why detectors transfer across models sharing the same deconvolution structure.","feed_headline":"Strides in AI-music generators stamp telltale peaks into every song","feed_subtitle":"A Fourier proof shows the artifact's positions are fixed by stride sizes, enough for 99%-accuracy detection.","key_machinery":"The central mechanism is the Fourier transform's periodization identity: multiplying a signal by a Dirac comb samples it in time and copies its spectrum at intervals in frequency. The paper reinterprets zero-insertion upsampling as such a comb multiplication, so each deconvolution layer tiles copies of its input spectrum. The artifact fingerprint, defined as the average spectrum after subtracting local minima over a sliding window and restricted to a high-frequency band, is the operational object used to read the resulting peaks. It is the quantity the linear detector acts on.","core_discovery":"Stated on the paper's own terms: a strided deconvolution is a zero-upsampling followed by a 1-strided convolution; in Fourier terms the zero-upsampling periodizes the spectrum, replicating the input spectrum at copies separated by the original sampling frequency $f_s$. Because biases and ReLU activations concentrate energy at $0$ Hz, the periodized copies appear as peaks at frequencies $n f_s$ for $n \\leq \\lfloor k/2 \\rfloor$, up to the new Nyquist frequency $k f_s/2$. Through several deconvolution layers, these peaks clone recursively, so the final output contains $\\lfloor P_{\\max}/2\\rfloor+1$ peaks where $P_{\\max}=\\prod_i k^{(i)}$ is the product of the strides; Encodec's strides $\\{8,5,4,2\\}$ yield 161 predicted peaks. The paper validates this with experiments on DAC, Encodec, and Musika!, and on Suno and Udio, and shows that the peak locations are the same for the same architecture trained on different datasets with different random seeds.","pith_inferences":["If this account is right, a model that avoids detectable peaks would have to replace deconvolutions with other upsampling methods or learn kernels whose frequency response vanishes at the replicated peak positions—a concrete architectural target for future generators.","An adversary could in principle spoof the fingerprint by injecting matching peaks into real audio, which suggests that peak-presence alone may weaken as a detector over time and that phase or higher-order statistics may become necessary.","The failure on the unseen Udio 32 version suggests zero-shot detector transfer is really an architecture-matching problem; a testable prediction is that transfer will succeed exactly when two generators share the same deconvolution strides and fail when they do not."],"forward_implications":["Detection is possible from a few engineered features: a logistic regression over the artifact fingerprint gives roughly 99% accuracy on open-source codecs and on commercial generators such as Suno v3.5 and Udio 130.","A detector trained on one model or dataset transfers to another model that shares the same deconvolution strides, because the peak positions are determined by the architecture alone.","The artifact pattern encodes the stride hyperparameters of the generator, so the recursive peak count can be read off as an architecture fingerprint.","The artifact cannot be removed by better training data or longer training; it is inherent to the use of deconvolution layers, and only an architectural change or carefully chosen kernels could eliminate it.","The Fourier analysis applies to the deconvolution layer itself, so the phenomenon is not specific to music: any generative model built on deconvolutions is predicted to carry a similar spectral fingerprint in its output."],"supporting_citations":[{"why":"Supplies the original checkerboard-artifact analysis and the view of deconvolution as upsampling plus convolution, the phenomenon this paper re-derives in Fourier terms.","marker":"[15]"},{"why":"Provides the formal rewriting of a strided deconvolution as zero-insertion upsampling followed by a 1-strided convolution, which is the basis for the periodization proof.","marker":"[32]"},{"why":"Supplies the Fourier transform properties used in the proof, including multiplication-convolution duality and spectral interpolation.","marker":"[31]"},{"why":"The Encodec model whose decoder stride set {8,5,4,2} yields the 161-peak prediction and is used in the experiments.","marker":"[33]"},{"why":"DAC, the neural codec used for the architecture-dependence experiments and as a detection benchmark.","marker":"[34]"},{"why":"The nearest detection baseline, providing the experimental setup, codec choices, FMA medium, and the intra-family generalization result this paper compares against.","marker":"[17]"},{"why":"The SONICS dataset, used for training and evaluating the detector on Suno and Udio, including the unseen-version splits.","marker":"[16]"},{"why":"The average-spectrum artifact extraction idea from computer vision that motivates the paper's artifact fingerprint processing.","marker":"[27]"}],"fun_headline_variants":["Fourier proof: AI music's telltale peaks are stride-imposed","Stride sizes set AI music's spectral fingerprints","AI music artifacts traced to mathematical stride effect","AI music's stride peaks enable 99% detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that real generators' hidden layers have a strong low-frequency or DC component that survives both the learned convolution kernels and the nonlinearities; if a generator learned kernels that cancel the periodized peaks, the artifact and the detector would disappear.","fun_headline_variants_meta":{"raw":{"variants":["Fourier proof: AI music's telltale peaks are stride-imposed","Stride sizes set AI music's spectral fingerprints","AI music artifacts traced to mathematical stride effect","AI music's stride peaks enable 99% detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000651,"raw_usage":{"total_tokens":3001,"prompt_tokens":978,"completion_tokens":2023,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":1957}},"tokens_in":594,"tokens_out":2023,"duration_ms":14278,"temperature":1.0,"reasoning_tokens":1957,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:37:32.350375+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct or train an audio generator whose deconvolution kernels are constant over the stride length, so their frequency response cancels the replicated peaks, while hidden layers are kept free of a DC bias; if the predicted peaks at $n f_s$ still appear in the average spectrum, the theory as stated is incomplete, and if they vanish, the artifact is not strictly unavoidable.","supporting_citations":[{"cited_title":"Slopa- ganda: The interaction between propaganda and gener- ative AI,","cited_arxiv_id":null,"evidence_quote":"Supplies the original checkerboard-artifact analysis and the view of deconvolution as upsampling plus convolution, the phenomenon this paper re-derives in Fourier terms."},{"cited_title":"Intriguing properties of synthetic im- ages: from generative adversarial networks to diffusion models,","cited_arxiv_id":null,"evidence_quote":"Provides the formal rewriting of a strided deconvolution as zero-insertion upsampling followed by a 1-strided convolution, which is the basis for the periodization proof."},{"cited_title":"Detecting generated images by real images,","cited_arxiv_id":null,"evidence_quote":"Supplies the Fourier transform properties used in the proof, including multiplication-convolution duality and spectral interpolation."},{"cited_title":"Adver- sarial audio synthesis,","cited_arxiv_id":null,"evidence_quote":"The Encodec model whose decoder stride set {8,5,4,2} yields the 161-peak prediction and is used in the experiments."},{"cited_title":"Melgan: Generative adversarial networks for conditional waveform synthesis,","cited_arxiv_id":null,"evidence_quote":"DAC, the neural codec used for the architecture-dependence experiments and as a detection benchmark."},{"cited_title":"AI Art is Theft: Labour, Extraction, and Exploitation: Or, On the Dangers of Stochastic Pol- locks,","cited_arxiv_id":null,"evidence_quote":"The nearest detection baseline, providing the experimental setup, codec choices, FMA medium, and the intra-family generalization result this paper compares against."},{"cited_title":"The creation and detection of deepfakes: A survey,","cited_arxiv_id":null,"evidence_quote":"The average-spectrum artifact extraction idea from computer vision that motivates the paper's artifact fingerprint processing."}],"review_version":2}