{"id":"e151176c-c2ac-4e46-b7c1-5a7953c65cda","arxiv_id":"1908.08813","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A fast Capon-based ENF estimator with temporal windowing is proposed, and window selection is shown to matter more than estimator sophistication for accuracy.","lead":"This paper tests a faster way to measure the electric network frequency (ENF) from power and speech recordings, using a Capon spectral estimator with temporal windowing. It finds that choosing the right window, like Parzen, can make even simple STFT methods outperform complex state-of-the-art ENF estimators.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Gohberg-Semencul fast inversion in Sec. 3.2 is unjustified for windowed data: \\hat R in Eq. (5) is not Toeplitz once a non-rectangular window is applied, and the notation shifts from the (m+1)-dimensional \\hat R to an N-dimensional R_N, so the fast algorithm's target matrix is undefined.","rationale":"The reader and I converge on the Toeplitz gap as the central weakness. The strongest claim is a fast Capon method that outperforms state-of-the-art ENF estimators; for that claim to hold, the fast inversion must invert the matrix that Eq. (6) actually uses. The temporal window makes \\hat R non-Toeplitz, and the paper does not demonstrate the low-displacement-rank property that would justify Eq. (9). The dimension shift from an (m+1)-dimensional \\hat R to an N-dimensional R_N makes it unclear what was computed. I am not objecting to the empirical observation that temporal windowing improves STFT/periodogram ENF estimates; that is a plausible and interesting result supported by concrete correlations. But the proposed fast implementation is the paper's technical contribution, and its derivation is not merely missing polish: no standard reading of Eqs. (5)-(13) yields the claimed algorithm. A direct-versus-fast comparison is cheap and would settle whether the reported correlations are genuine Capon values or artifacts of an approximate or undefined inversion. Since the reader's REJECT and high correctness risk match my reading, I recommend no change to the verdict.","tokens_in":8575,"tokens_out":7707,"duration_ms":77229,"concrete_test":"Re-run the Data 1 evaluation for 1-second frames in two variants: (A) direct Cholesky inversion of the (m+1)x(m+1) \\hat R from Eq. (5), then Eq. (6); (B) the GS/Krylov procedure in Sec. 3.2. Compare the ENF correlation coefficients against Table 1. If (A) and (B) differ beyond rounding, the fast inversion is not computing the Capon estimate it claims. If they match, verify on a synthetic WSS sinusoid with a Hamming window that the GS inverse actually equals the direct inverse of \\hat R; a mismatch would confirm the Toeplitz assumption is violated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (5) forms \\hat R = 1/(N-m) \\sum_{t=m+1}^N [\\tilde y(t),...,\\tilde y(t-m)][\\tilde y^*(t),...,\\tilde y^*(t-m)] with \\tilde y(t)=w(t-n)y(t) and m=10. Its (i,j) entry is proportional to \\sum_t w(t-i-n)w(t-j-n)y(t-i)y^*(t-j); even in expectation over a WSS y this depends on the window product w(t-i)w(t-j), not only on i-j. Thus \\hat R is not Toeplitz for the Parzen, Hamming, or Kaiser windows used in the paper. The Gohberg-Semencul factorization in Eq. (9) reconstructs the inverse from first-column quantities via Eqs. (10)-(13) and requires Toeplitz or low-displacement-rank structure; no proof or approximation is supplied for the windowed sample covariance. Moreover, Eqs. (5)-(6) use an (m+1)x(m+1) covariance with m=10, whereas Sec. 3.2 phrases the Krylov/GS inversion for an N x N R_N with N=LFs. Since Eq. (6) needs only the inverse of the 11x11 matrix, the fast factorization is either operating on a different matrix or the Capon formula is misstated. The claimed efficiency and the reported correlation values rest on this step, so the central algorithmic contribution is not supported as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a fast Capon-based spectral estimator for Electric Network Frequency (ENF) extraction, combining temporal windowing with Gohberg–Semencul factorization and Krylov matrices to speed up matrix inversion. Experiments on power and speech recordings report high correlation with reference ENF signals and claim improvements over several state-of-the-art methods, especially at short frame lengths. The paper also investigates the effect of different temporal windows on both the proposed method and the STFT baseline.","tokens_in":8971,"tokens_out":4265,"duration_ms":40013,"significance":"If the algorithmic claims held, the method would be a practical contribution to ENF-based multimedia forensics, offering high accuracy with low computational cost at short frame lengths. The paper includes extensive comparisons with multiple baselines and attempts a statistical significance test. However, the central derivation is not sound: the fast inversion relies on a Toeplitz property that does not hold for the windowed sample covariance matrix, so the main claim of an efficient Capon-based method is unsupported as written.","major_comments":[{"comment":"The fast implementation applies the Gohberg–Semencul (GS) factorization to the sample covariance matrix R-hat of Eq. (5), but after multiplication by a non-rectangular temporal window, R-hat is not Toeplitz. For windowed samples y-tilde(t) = w(t-n)y(t), the (i,j) entry of R-hat is proportional to sum_t w(t-i)w(t-j)y(t-i)y*(t-j), which depends on i and j through the window product, not only on i-j. The GS factorization requires Toeplitz or low-displacement-rank structure, and the paper supplies neither a proof nor an approximation showing that the windowed sample covariance satisfies this condition. This is the load-bearing step for the claimed computational efficiency and for the reported accuracy, so the central contribution is not established.","section":"Section 3.2 and Eq. (5)"},{"comment":"There is a dimensional inconsistency between the covariance matrix used in the Capon formula and the one inverted by the GS algorithm. Eq. (5) defines an (m+1)x(m+1) matrix with m=10, and Eq. (6) only needs the inverse of this small matrix. In contrast, Eqs. (9)–(13) describe a fast inversion of an NxN matrix R_N, with N = LFs. The paper does not clarify how the N-dimensional matrix relates to the (m+1)-dimensional sample covariance, so either the fast algorithm is operating on a different matrix than the one used in Eq. (6), or the Capon formula is misstated. The claimed complexity reduction is therefore not backed by a coherent algorithmic description.","section":"Section 3.2 vs. Section 3.1"},{"comment":"The harmonic index and temporal window are selected based on performance on the same test recordings (e.g., 'only the third harmonic was used, because it provides the best results' for Data 1, Section 4.1). This in-sample selection inflates the reported correlation coefficients and makes the comparisons to state-of-the-art methods optimistic. The authors should separate model selection from evaluation, for example by using a validation subset or cross-validation.","section":"Section 4.1 and Tables 1–4"},{"comment":"The significance test assumes n=1800 independent samples, but the ENF sequences being correlated are estimated from frames with a 1-second shift and frame lengths up to 20 seconds, so successive estimates are highly dependent. Moreover, the pairwise correlations are computed on the same recordings, so the samples are not independent across methods either. The statement that all pairwise differences are significant at the 95% level is therefore not statistically justified.","section":"Section 4.2, Fisher transform"}],"minor_comments":[{"comment":"'Kaizer' is a typo for 'Kaiser'; please correct throughout the manuscript.","section":"Table 2 and Table 4"},{"comment":"The reference to 'Fig. 1' in the text does not appear to be accompanied by the actual figure or its data in the manuscript; include the figure or remove the reference.","section":"Section 4.2"},{"comment":"The notation g-tilde(l) is used in the correlation formula but is not explicitly defined in the equation; define it for clarity.","section":"Eq. (3)"},{"comment":"The claim that this is the first use of a non-rectangular temporal window in ENF estimation should be verified against the broader ENF literature, which includes many windowing practices, or softened to avoid overstatement.","section":"Abstract and Section 2"}],"recommendation":"reject","confidential_remarks":"The core theoretical problem in Section 3.2 is a misapplication of the GS factorization to a matrix that lacks the required structure; this is not a minor omission. The experimental protocol also involves in-sample selection of windows and harmonics. I do not see a simple fix within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick take on 1908.08813. The one thing worth remembering is that temporal windowing before spectral estimation helps ENF extraction, and STFT with a Parzen window gets surprisingly close to state-of-the-art at 1-second frames. That is a nice empirical point and, as far as I know, new to the ENF literature. The Capon-with-windowing idea is also reasonable, and the paper does a broad comparison across frame lengths and windows on two datasets. The experimental design is standard for the field, and the reference list looks appropriate, including their own prior work where relevant.\n\nThe soft spot is exactly where the stress-test note points. The paper claims to use Gohberg-Semencul factorization because the covariance matrix is Toeplitz, but after multiplying the frame by a non-rectangular window the sample covariance in Eq. (5) is not Toeplitz, even in expectation. The (i,j) entry contains w(t-i)w(t-j), which depends on absolute position, not just lag. The fast inversion in Sec. 3.2 has no stated justification for this matrix. There is also a notation shift: Eq. (5) defines an 11x11 \\hat R with m=10, while Sec. 3.2 talks about an N x N R_N. So the target matrix for the GS factorization is undefined. That is a load-bearing flaw for the \"efficient\" part of the title. That said, m=10 is small, so a direct inversion would be cheap; the empirical numbers could be reproducible with a correct implementation. The paper just doesn't describe that.\n\nOther issues are minor in comparison: window and harmonic choices are made on the same test data, so there is in-sample selection; the correlation coefficients come without error bars; the Fisher z-test treats dependent samples as independent; and the headline STFT result appears only in a figure, not a number-laden table. None of these are fatal on their own.\n\nWho is this for? People working on ENF extraction, especially forensic applications with short audio clips, will find the windowing observation useful. The method section needs fixing before anyone should rely on the Capon speed claims.\n\nRecommendation: this deserves a serious referee, not a desk reject, because the empirical result is real and the fix (drop the GS framing, or justify it for a different matrix) is likely within reach. As written, it's a weak reject. If I were editing, I'd send it out with major revision and see if the authors can clean up the algorithm.","headline":"The temporal-windowing observation is real and useful for ENF extraction, but the fast-Capon derivation has an unjustified Toeplitz assumption and a notation mismatch that sink the efficiency claim as written.","tokens_in":9448,"tokens_out":4515,"would_cite":false,"duration_ms":46208,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Applying a non-rectangular temporal window before spectral estimation lets a fast Capon-based estimator recover the Electric Network Frequency from one-second power recordings with correlation 0.9990, and even plain STFT with a Parzen…","keywords":["Electric Network Frequency","Capon spectral estimator","temporal windowing","Gohberg-Semencul factorization","Krylov matrices","Parzen window","multimedia forensics","frequency estimation"],"falsifier":"Take a real one-second power-recording frame, build $\\hat{R}$ exactly as in Eq. (5) with a Parzen window, and test whether $\\hat{R}$ is Toeplitz or has the assumed displacement rank. Then compare the Gohberg-Semencul inverse from Eq. (9) with a direct inverse: if the matrices disagree materially, or if replacing the fast inversion with a direct inversion changes the reported 0.9990 correlation or runtime, the fast-implementation claim fails.","tokens_in":8369,"feed_emoji":"⚡","tokens_out":7980,"duration_ms":72425,"temperature":0.7,"pith_summary":"The paper tries to establish that temporal windowing, applied before spectral estimation, is a decisive and previously underused step in extracting the Electric Network Frequency (ENF) from audio and power recordings. It proposes a fast Capon-based spectral estimator that multiplies each frame by a non-rectangular temporal window—Parzen works best for power recordings—and speeds up covariance inversion through a Gohberg-Semencul factorization. On a 30-minute power recording, the method reaches a correlation of 0.9990 with the reference ENF even with a one-second frame, where competing methods fall below 0.99. The paper also reports that plain Short-Time Fourier Transform with a Parzen window reaches about 0.9990, outperforming recent state-of-the-art estimators. This matters because forensic authentication often must work on short recordings, and speed matters for large datasets.","feed_headline":"Temporal windowing lifts ENF accuracy to 0.9990 on 1-sec clips","feed_subtitle":"Fast Capon with a Parzen window tracks the power-grid frequency from one-second clips better than recent ENF estimators; even plain STFT…","key_machinery":"The load-bearing object is the filter-bank Capon spectral estimator combined with temporal windowing. Capon uses a data-dependent finite-impulse-response filter that passes frequency $\\omega$ while minimizing output power from all other frequencies, giving the high-resolution spectral estimate $\\hat{\\varphi}(\\omega) = (m+1)/(a^*(\\omega)\\hat{R}^{-1}a(\\omega))$. The speed-up relies on writing $\\hat{R}^{-1}$ via Krylov matrices and the Gohberg-Semencul factorization (Eq. 9), an exact expression of the inverse of a Toeplitz matrix as a difference of products of Krylov matrices, so the denominator at all frequency bins is evaluated with one zero-padded FFT instead of repeated matrix inversions. The temporal window—Parzen in the main power experiments—is the conceptual novelty: it is applied to the time series before covariance estimation, and the paper argues this choice, not the spectral estimator alone, controls estimation accuracy at short frame lengths.","core_discovery":"The authors claim that a filter-bank Capon spectral estimator, fed with temporally windowed frames, gives higher correlation with the ground-truth ENF than the state-of-the-art methods they compare against, while remaining computationally fast. The estimator forms the sample covariance $\\hat{R}$ from windowed samples $\\tilde{y}(t) = w(t-n)y(t)$, computes the data-dependent Capon filter $h = \\hat{R}^{-1}a(\\omega)/(a^*(\\omega)\\hat{R}^{-1}a(\\omega))$, and evaluates the spectral denominator through a zero-padded FFT after expressing $\\hat{R}^{-1}$ with Krylov matrices and the Gohberg-Semencul factorization, Eq. (9). With a Parzen window and a 1-second frame, the estimated ENF correlates at 0.9990 with the ground truth; at 5 and 10 seconds it reaches 0.9991. On the speech dataset, a rectangular window gives 0.9351 at 33 seconds, close to linear prediction but faster. The authors further show that window selection is decisive: Parzen or Hamming windows are needed for short power-record frames, while Kaiser and rectangular windows fail below 10 seconds, and the same window logic makes even STFT outperform recent specialized estimators.","pith_inferences":["An extension the paper leaves implicit is jointly optimizing the window shape and frame length per harmonic's signal-to-noise ratio, which could push ENF accuracy further and could be tested on the same datasets.","The reported gap between windowed STFT and windowed Capon is small, raising the question of whether the Capon estimator's advantage persists once both methods use the same optimal window on a wider range of recordings.","A natural extension, which the conclusion gestures toward, is applying the windowed fast-Capon pipeline to single images with rolling-shutter or fluorescent-light ENF signatures, where short analysis windows are the norm."],"forward_implications":["Forensic timestamping and authentication can work on power recordings as short as one second, since the estimated ENF matches the reference at 0.9990 correlation.","Window selection should become a standard control variable in ENF studies; the paper shows a Parzen-windowed STFT beating recent specialized estimators, so previous comparisons that used only rectangular windows may need revisiting.","The fast Capon implementation makes high-resolution spectral estimation practical for large speech datasets, where speed currently favors simpler methods.","For speech recordings, where the rectangular window wins, the proposed method yields 0.9351 correlation at 33 seconds, close to linear prediction but with lower time cost."],"supporting_citations":[{"why":"Supplies the adaptive ENF extraction pipeline, quadratic interpolation, and correlation-based matching that the proposed method adopts and compares against.","marker":"[3]"},{"why":"Provides the maximum-likelihood multitone harmonic ENF estimator used as a baseline in Tables 1 and 3.","marker":"[4]"},{"why":"Provides the rank-reduction linear-prediction ENF estimates that the proposed method is claimed to outperform in correlation and speed.","marker":"[7]"},{"why":"Supplies the systematic comparison of spectral estimation methods and the parametrization the authors follow for filtering and frame processing.","marker":"[8]"},{"why":"Supplies the Gohberg-Semencul fast inversion of the covariance matrix via Krylov matrices, the core of the proposed fast implementation.","marker":"[19]"},{"why":"Supplies the efficient iterative-adaptive-approach implementation and the linear system used to compute the factorization parameters.","marker":"[20]"},{"why":"Defines the Capon filter-bank spectral estimator and its covariance model, which the proposed method adapts.","marker":"[21]"},{"why":"Defines the Parzen spectral window used in the main power-recording experiments.","marker":"[22]"},{"why":"Provides the window-function taxonomy, including Parzen and Hamming windows, used for the temporal windowing experiments.","marker":"[23]"}],"fun_headline_variants":["Capon with windowing hits 0.999 ENF accuracy in 1 sec","Windowed Capon beats prior ENF estimates in under a second","Temporal window makes even STFT beat top ENF estimators","Fast ENF: windowed Capon outdoes newer methods on 1-sec clips","Windowed Capon: 0.9990 ENF correlation in one second"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speed-up assumes the covariance matrix built from windowed frames has a shift-invariant (Toeplitz) or low-displacement-rank structure; multiplying the frame by a non-rectangular window generally destroys that structure for the sample covariance in Eq. (5), and the paper provides no argument or approximation showing the Gohberg-Semencul factorization still applies.","fun_headline_variants_meta":{"raw":{"variants":["Capon with windowing hits 0.999 ENF accuracy in 1 sec","Windowed Capon beats prior ENF estimates in under a second","Temporal window makes even STFT beat top ENF estimators","Fast ENF: windowed Capon outdoes newer methods on 1-sec clips","Windowed Capon: 0.9990 ENF correlation in one second"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000746,"raw_usage":{"total_tokens":3369,"prompt_tokens":1031,"completion_tokens":2338,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":2236}},"tokens_in":647,"tokens_out":2338,"duration_ms":15678,"temperature":1.0,"reasoning_tokens":2236,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:43:57.161019+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real one-second power-recording frame, build $\\hat{R}$ exactly as in Eq. (5) with a Parzen window, and test whether $\\hat{R}$ is Toeplitz or has the assumed displacement rank. Then compare the Gohberg-Semencul inverse from Eq. (9) with a direct inverse: if the matrices disagree materially, or if replacing the fast inversion with a direct inversion changes the reported 0.9990 correlation or runtime, the fast-implementation claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive ENF extraction pipeline, quadratic interpolation, and correlation-based matching that the proposed method adopts and compares against."},{"cited_title":"Datasets Two datasets are employed in order to evaluate the proposed approach and compare it with other state-of-the-art methods","cited_arxiv_id":null,"evidence_quote":"Provides the maximum-likelihood multitone harmonic ENF estimator used as a baseline in Tables 1 and 3."},{"cited_title":"Applications of ENF criterion in forensic audio, video, computer and telecommunication analy- sis,","cited_arxiv_id":null,"evidence_quote":"Provides the rank-reduction linear-prediction ENF estimates that the proposed method is claimed to outperform in correlation and speed."},{"cited_title":"ENF extrac- tion from digital recordings using adaptive techniques and frequency tracking,","cited_arxiv_id":null,"evidence_quote":"Supplies the systematic comparison of spectral estimation methods and the parametrization the authors follow for filtering and frame processing."},{"cited_title":"Edit detection in speech recordings via instantaneous electric network frequency variations,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gohberg-Semencul fast inversion of the covariance matrix via Krylov matrices, the core of the proposed fast implementation."},{"cited_title":"Audio authenticity: Detecting ENF disconti- nuity with high precision phase analysis,","cited_arxiv_id":null,"evidence_quote":"Supplies the efficient iterative-adaptive-approach implementation and the linear system used to compute the factorization parameters."},{"cited_title":"Au- dio authentication by exploring the absolute-error-map of ENF signals,","cited_arxiv_id":null,"evidence_quote":"Defines the Capon filter-bank spectral estimator and its covariance model, which the proposed method adapts."},{"cited_title":"Audio authentication using the kurtosis of esprit based ENF estimates,","cited_arxiv_id":null,"evidence_quote":"Defines the Parzen spectral window used in the main power-recording experiments."},{"cited_title":"Factors affecting ENF capture in audio,","cited_arxiv_id":null,"evidence_quote":"Provides the window-function taxonomy, including Parzen and Hamming windows, used for the temporal windowing experiments."}],"review_version":1}