{"id":"029810ff-b2a8-4e15-ab9e-c0bcad7213df","arxiv_id":"2412.08856","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"SEDM couples separate magnitude and phase diffusion networks with real-noise corruption and a cycle-consistency loss, reporting improved speech-enhancement metrics on IEEE, TIMIT, VCTK, and DNS test sets.","lead":"Researchers built a speech enhancement model that cleans noisy audio by separately estimating the loudness (magnitude) and timing offset (phase) of each frequency, using two linked diffusion networks. They report better scores than several published systems on standard benchmarks by also mixing in real noise clips during training and cross-checking magnitude against phase.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reverse update Eq. (4) depends on a noise term I_m whose test-time provenance is never defined or trained; as written, SEDM is underspecified and its reported SOTA numbers cannot be reproduced from the paper.","rationale":"The reviewer's weakest assumption identifies the test-time I_m, which is the same core gap I focus on. This is the single most load-bearing issue because it blocks even a single forward pass at inference: without a rule for I_m, Eq. (4) cannot be evaluated, so the architecture as specified cannot produce the outputs in Tables 2 and 4. The paper's other weaknesses (unreleased code, missing equation numbers in Algorithm 1, no error bars) are real but secondary; they would affect confidence in numbers, not the definition of the method. An honest audit of the text shows no definition of I_m, so the central claim of SOTA improvement is not supported by a reproducible algorithm. The verdict stays REJECT; no adjustment is needed.","tokens_in":11862,"tokens_out":7639,"duration_ms":79334,"concrete_test":"Audit the paper for any equation or training loss that defines I_m in Eq. (4): specifically, search for a loss term like L_noise = ||I_m - I_n||² or any text stating how the noise estimate is computed at test time. If none is found, the reverse update is not a well-defined function of the network, confirming the concern. If one is found, reimplement the model with that definition and verify whether VCTK PESQ matches Table 2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim rests on the reverse process Eq. (4): Y_{m-1} = (1/γ_m)(Y_m - θ_m/(1-γ_m) I_m) + σ_m. The paper never defines I_m, nor does it specify a network output, a training loss, or an inference procedure for it. The text states 'the reverse network estimates both clean speech and noise spectra' (Section 'Reverse Process') but no equation supervises the noise estimate. In training, the forward process adds known noise clip I_n, but at test time the diffusion network is discarded; the only input is the noisy mixture. If I_m is meant to be the noise spectrogram of the mixture, it is not available; if it is meant to be an estimate, its objective is absent. Thus Eq. (4) is not computable from the paper's description, making the method underspecified. Without this term, the reverse update cannot be applied to test utterances; the benchmark tables (Tables 2 and 4) therefore cannot be attributed to the described algorithm. This is not a matter of tuning or baseline fairness; it is a missing definition in the core generative step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SEDM, a diffusion-model speech enhancement method that separates magnitude and phase estimation into two networks, replaces Gaussian noise with real-world noise clips in the forward process, and uses a complex-cycle-consistent (CCC) loss to couple magnitude and phase estimates. The authors report results on IEEE, TIMIT, VCTK, and DNS challenge benchmarks, claiming state-of-the-art performance, e.g., SEDM-L reaching PESQ 3.59 on VCTK and outperforming NASE by 0.58 PESQ and 0.04 ESTOI.","tokens_in":12101,"tokens_out":3845,"duration_ms":39735,"significance":"If the described algorithm were fully specified and reproducible, the reported gains would be significant for monaural speech enhancement, particularly the combination of separate magnitude/phase diffusion and a cycle-consistent coupling mechanism. The paper includes a broad set of benchmarks, an ablation study, and a claim of fair baseline reproduction for the non-diffusion competitors, which are strengths. However, the central generative step is currently underspecified, and the reported numbers cannot be assigned to the described algorithm without additional definitions.","major_comments":[{"comment":"The reverse update in Eq. (4) depends on a term I_m whose definition and test-time provenance are never given. The text states that 'the reverse network estimates both clean speech and noise spectra,' but no network output, training loss, or inference procedure is specified for I_m; during test the diffusion network is discarded and the only input is the noisy mixture. Without a precise definition of I_m (e.g., an estimated noise spectrogram with a supervised loss), Eq. (4) is not computable and the benchmark results in Tables 2 and 4 cannot be attributed to the described algorithm. This is the central load-bearing gap.","section":"Reverse Process, Eqs. (4)-(5)"},{"comment":"The forward process is written as an SDE dY_n = μ(Y_n,n)dn + σ(Y_n,n)dI_n, where dI_n is the increment of a randomly selected real-noise clip rather than a Wiener process; no stochastic integral or discrete transition q(Y_n|Y_{n-1}) is defined. Consequently, the claimed forward process is not a standard diffusion and the reverse update is not derived from it. The scheduling parameters γ_m, θ_m and the quantities \\barγ_m appearing in Eq. (5) are also undefined. Please replace Eq. (3) with an explicit discrete forward transition and derive the reverse posterior, or state clearly that Eq. (4) is a heuristic update rather than the exact reverse of Eq. (3).","section":"Diffusion Process, Eqs. (2)-(3)"},{"comment":"The CCC loss definitions contain internal inconsistencies that prevent reproduction. The loss L_{SA} in Eq. (7) is never defined; the text says the magnitude loss is estimated 'with (3),' but Eq. (3) is the forward SDE. Algorithm 1 line 10 says 'Update L_{S'_A} as (8),' but Eq. (8) defines the phase reconstruction loss L_{SA→P}, and line 12 refers to a nonexistent Eq. (10). Please renumber the equations and align Algorithm 1 with the displayed loss definitions.","section":"Complex-Cycle-Consistent Learning, Eqs. (6)-(9) and Algorithm 1"},{"comment":"For the diffusion-model comparison in Table 4, the paper does not state whether the baseline numbers were reproduced under the same training data, SNR conditions, and evaluation pipeline as the proposed SEDM, unlike the statement made for the competitors in Section 'Competitors.' Since the claimed 0.58 PESQ gain over NASE is a headline result, please specify the provenance of each baseline entry and confirm that the settings match Table 2.","section":"Comparison to Other Diffusion Models, Table 4"}],"minor_comments":[{"comment":"The table captions are inconsistent: Table 6 is referred to as 'Table 1' in the text and Table 7 is referred to as 'Table 2,' and the heading 'Benckmarks on TIMIT' is a typo.","section":"Quantitative Results"},{"comment":"The caption says 'left side of Figure 1' for both the diffusion and reverse networks; the reverse network should be on the right side.","section":"Figure 1 caption"},{"comment":"The noise type names 'dwashing,' 'dliving,' and 'pstation' appear to be typos for the DEMAND classes 'dwashing,' 'dliving,' and 'pstation.'","section":"Introduction"},{"comment":"The symbol S_{t,f} is used for the complex STFT in Eq. (1), while S_A and S_P later denote magnitude and phase spectra; please define these symbols consistently.","section":"Eq. (1) and CCC notation"},{"comment":"The baseline 'ResNet152' is mentioned, but the input and output of this non-diffusion baseline are not described; please clarify how it is used.","section":"Ablation Study"}],"recommendation":"major_revision","confidential_remarks":"The missing definition of I_m in Eq. (4) is the main correctness risk, and the stress-test concern is valid. I recommend major revision rather than rejection because the gap is potentially fixable by adding a derivation, a training loss for the noise estimate, and a test-time procedure. If the authors cannot provide those missing elements, the paper should not be accepted. The paper also shows signs of an unfinished submission (broken cross-references, duplicate reference for Welker et al., inconsistent table numbering), which the authors should fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper combines several existing ideas in a way that is genuinely new for speech enhancement: separate magnitude and phase diffusion networks, real-world noise clips in the forward process, and a cycle-consistent loss between magnitude and phase. The benchmark numbers are strong (e.g., PESQ 3.59 on VCTK, beating SCP-GAN by 0.07) and the ablation study is a sensible attempt to credit each contribution. If the numbers are accurate, this is a step forward for the subfield.\n\nThe problem is that the method as described cannot be reproduced. The reverse update in Eq. (4) depends on a noise term I_m, but the paper never says where I_m comes from at test time. The reverse network is said to 'estimate both clean speech and noise spectra,' yet no loss supervises the noise estimate and no procedure feeds it into Eq. (4). This is not a minor typo; it is the core generative step, so the reported tables cannot be attributed to the described algorithm. The SDE in Eq. (3) is also loosely written: dI_n is not a standard increment, and the stochastic integral is undefined. Add to that the mis-referenced equations in Algorithm 1 and the absence of code or data splits, and a reader cannot verify anything.\n\nThe empirical claims are also a bit oversold. On the DNS Challenge, SEDM-L does not beat RemixIT on WP or SI-SDR; it wins only on NP and STOI. The text acknowledges this as 'competitive,' but the abstract and conclusion claim general superiority. The gains over the strongest baselines are modest, and there are no error bars, so it is hard to know if the differences are meaningful.\n\nThat said, the core idea is clear and potentially useful. The missing details are fixable: define I_m properly, specify the training loss for the noise estimate, clean up the SDE, release code or a detailed derivation. This is exactly the kind of paper that should go to peer review, not be desk-rejected, because the conceptual contribution is solid and the flaws are mostly in presentation and specification. A careful referee could push the authors to make the method reproducible.\n\nWho gets value from this: researchers working on diffusion-based speech enhancement, and anyone thinking about phase-magnitude coupling in generative models. I would not cite it in its current form, but I would read a revised version. Send it to review with an expectation of major revision.","headline":"A promising diffusion-based speech enhancement recipe with a genuinely new magnitude-phase cycle-consistency idea, but the reverse process is underspecified around a load-bearing noise term, and the paper is not reproducible as written.","tokens_in":12659,"tokens_out":2851,"would_cite":false,"duration_ms":32167,"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":"SEDM couples magnitude and phase diffusion with a cycle-consistent loss and reports new best scores on VCTK and DNS benchmarks.","keywords":["speech enhancement","diffusion model","cycle-consistent learning","phase estimation","magnitude and phase","real-world noise","monaural speech enhancement","spectrogram"],"falsifier":"A direct test would be to run SEDM-L on VCTK test utterances while replacing the estimated noise spectrum in the reverse update with the true noise clip, and again with a zero or mismatched noise spectrum; if the PESQ difference between these runs is large, the method is silently relying on oracle noise information that the paper does not describe how to obtain at test time.","tokens_in":11570,"feed_emoji":"🎧","tokens_out":10842,"duration_ms":91728,"temperature":0.7,"pith_summary":"The paper tries to establish that a diffusion model for monaural speech enhancement works best when the magnitude and phase of the speech spectrogram are estimated by two separate networks, when the forward corruption uses real-world noise clips instead of Gaussian noise, and when a complex-cycle-consistent (CCC) loss forces the estimated magnitude and phase to map back and forth into each other. If this is right, the coupling of magnitude and phase is an important source of gains: the largest model, SEDM-L, reports PESQ 3.59 on VCTK, beating the previous best diffusion method NASE by 0.58 PESQ and 0.04 ESTOI. The result is a concrete recipe for phase-aware enhancement that other systems could adopt.","feed_headline":"Diffusion model lifts speech-enhancement PESQ to 3.59 on VCTK","feed_subtitle":"Separate magnitude and phase networks plus a cycle-consistency loss beat the best prior diffusion method.","key_machinery":"The load-bearing machinery is the pair of separate diffusion paths for magnitude and phase plus the Complex-Cycle-Consistent (CCC) block. In the forward pass, real-world noise clips selected from the DEMAND database are progressively added to the clean magnitude and phase spectra through an SDE. In the noise-aware reverse pass, each step obeys $Y_{m-1} = \\frac{1}{\\gamma_m}(Y_m - \\frac{\\theta_m}{1-\\gamma_m} I_m) + \\sigma_m$, where $I_m$ is an estimated noise spectrum, and the two reverse branches output estimated magnitude $S_A$ and phase $S_P$. The CCC block then cycles $S_A$ through a mapping to phase and $S_P$ through a mapping to magnitude, adding cycle losses to the direct L2 losses so that each modality is trained to be consistent with the other. This cycle is the mechanism the paper claims exploits the intrinsic magnitude-phase relationship.","core_discovery":"On the paper's own terms, the central discovery is that phase and magnitude in the short-time Fourier transform are not independent targets: a cyclic mapping between them acts as a training signal that sharpens both estimates. The paper implements this in SEDM, where a forward process injects real-world noise clips into clean magnitude and phase spectra and a noise-aware reverse process estimates both clean and noise spectra at every step. Two LSTM-based cycle-consistency blocks then translate estimated magnitude to phase and estimated phase to magnitude, and the reconstruction errors are added to the direct estimation losses. With 50 diffusion blocks, SEDM-L reaches PESQ 3.59, CSIG 4.06, CBAK 4.22, COVL 4.89, and SSNR 11.8 on VCTK, and 2.93 wide-band PESQ on the DNS challenge, outperforming the compared state-of-the-art methods on most metrics. Ablations attribute the gain to the combination of the diffusion model, phase-aware estimation, and CCC, with the full model reaching PESQ 2.75 versus 2.21 without any of the three contributions.","pith_inferences":["A step the paper leaves implicit is an explicit loss or procedure for estimating the noise spectrum $I_m$ at test time; the current reverse update appears to require that spectrum without describing how it is supervised.","The CCC principle should transfer to other complex-spectrum regression tasks, such as source separation or dereverberation, wherever magnitude and phase are estimated jointly.","The comparison against Gaussian-trained models suggests that real-world noise corruption pays off most for deep networks; one testable extension is to measure how the gap scales with the number of diffusion blocks.","If the reverse update's noise spectrum is in fact an oracle at test time, the reported gains would shrink under a blind evaluation; a public implementation with a clearly specified noise-estimation path would settle this."],"forward_implications":["Following the SEDM recipe, with separate magnitude and phase diffusion branches, real-world noise corruption, and a CCC loss, yields the best reported scores among the compared methods on the VCTK and DEMAND setup, with SEDM-L at PESQ 3.59.","The ablation shows each ingredient contributes, with diffusion alone lifting PESQ from 2.21 to 2.43, phase-aware estimation adding further gains, and CCC bringing the full model to 2.75.","Diffusion models trained on real-world noise clips become competitive with Gaussian-trained models as network depth grows, and remain robust on seen noise while degrading on unseen noise types.","The CCC cycle makes magnitude and phase estimates mutually consistent, so the enhanced speech no longer has to borrow the phase of the noisy mixture."],"supporting_citations":[{"why":"Provides the DEMAND real-world noise database whose clips replace Gaussian noise in the forward diffusion process and in the test mixtures.","marker":"Thiemann, Ito, and Vincent 2013"},{"why":"DiffuSE is the diffusion-based speech enhancement baseline that introduced the supportive reverse process and is the main Gaussian-noise approach SEDM modifies.","marker":"Lu, Tsao, and Watanabe 2021"},{"why":"NASE is the noise-aware diffusion baseline that SEDM-L is compared against and outperforms by 0.58 PESQ and 0.04 ESTOI.","marker":"Hu et al. 2023"},{"why":"Establishes the intrinsic magnitude-phase relationship in the STFT that the CCC mechanism is designed to exploit.","marker":"Shimauchlt et al. 2017"},{"why":"Supplies the diffusion probabilistic model foundation that defines the forward and reverse processes used here.","marker":"Sohl-Dickstein et al. 2015"},{"why":"SGMSE is the score-based generative model in the complex STFT domain used as a comparison baseline.","marker":"Welker, Richter, and Gerkmann 2022b"},{"why":"Introduces complex ratio masking, the basis for treating magnitude and phase jointly in the complex domain.","marker":"Williamson, Wang, and Wang 2016"},{"why":"Defines the objective quality measures PESQ, CSIG, CBAK, and COVL used to evaluate the enhanced speech.","marker":"Hu and Loizou 2008"},{"why":"Demonstrates that realistic noise can boost diffusion-based denoising performance over Gaussian noise, motivating the same substitution in speech enhancement.","marker":"Wu et al. 2023"}],"fun_headline_variants":["Phase-magnitude cycle boosts diffusion speech enhancement to 3.59 PESQ","SEDM's cyclic phase-magnitude map hits PESQ 3.59 on VCTK","Cycle-consistent diffusion sharpens speech, PESQ 3.59 on VCTK","Diffusion learns phase-magnitude link for speech, PESQ 3.59","CCC diffusion: cycle magnitude-phase for speech PESQ 3.59"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reverse update formula requires the noise spectrum $I_m$ to be available or accurately estimated for the test utterance, yet the paper describes training with known noise clips and gives no loss or procedure that explains how the estimated noise is produced and fed into that equation at inference. If the test-time noise is not available in this form, the reverse process lacks a well-defined input and the reported comparisons would not transfer outside the training setup.","fun_headline_variants_meta":{"raw":{"variants":["Phase-magnitude cycle boosts diffusion speech enhancement to 3.59 PESQ","SEDM's cyclic phase-magnitude map hits PESQ 3.59 on VCTK","Cycle-consistent diffusion sharpens speech, PESQ 3.59 on VCTK","Diffusion learns phase-magnitude link for speech, PESQ 3.59","CCC diffusion: cycle magnitude-phase for speech PESQ 3.59"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000706,"raw_usage":{"total_tokens":3179,"prompt_tokens":942,"completion_tokens":2237,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":2120}},"tokens_in":558,"tokens_out":2237,"duration_ms":17358,"temperature":1.0,"reasoning_tokens":2120,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:29:27.929555+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to run SEDM-L on VCTK test utterances while replacing the estimated noise spectrum in the reverse update with the true noise clip, and again with a zero or mismatched noise spectrum; if the PESQ difference between these runs is large, the method is silently relying on oracle noise information that the paper does not describe how to obtain at test time.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DEMAND real-world noise database whose clips replace Gaussian noise in the forward diffusion process and in the test mixtures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the intrinsic magnitude-phase relationship in the STFT that the CCC mechanism is designed to exploit."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion probabilistic model foundation that defines the forward and reverse processes used here."},{"cited_title":"S.; Wang, Y.; and Wang, D","cited_arxiv_id":null,"evidence_quote":"Introduces complex ratio masking, the basis for treating magnitude and phase jointly in the complex domain."}],"review_version":1}